ℹ️ Skipped - page is already crawled
| Filter | Status | Condition | Details |
|---|---|---|---|
| HTTP status | PASS | download_http_code = 200 | HTTP 200 |
| Age cutoff | PASS | download_stamp > now() - 6 MONTH | 0 months ago |
| History drop | PASS | isNull(history_drop_reason) | No drop reason |
| Spam/ban | PASS | fh_dont_index != 1 AND ml_spam_score = 0 | ml_spam_score=0 |
| Canonical | PASS | meta_canonical IS NULL OR = '' OR = src_unparsed | Not set |
| Property | Value |
|---|---|
| URL | https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/ |
| Last Crawled | 2026-04-14 02:36:17 (10 hours ago) |
| First Indexed | 2023-07-10 13:30:24 (2 years ago) |
| HTTP Status Code | 200 |
| Meta Title | Black Sesame, Chinese auto chip challenger to Nvidia, burns $140M a year | TechCrunch |
| Meta Description | Black Sesame makes system-on-chips (SoCs) to power autonomous driving cars and other intelligent car functionalities. |
| Meta Canonical | null |
| Boilerpipe Text | Artificial intelligence is a deep and convoluted world. The scientists who work in this field often rely on jargon and lingo to explain what they’re working on. As a result, we frequently have to use those technical terms in our coverage of the artificial intelligence industry. That’s why we thought it would be helpful to put together a glossary with definitions of some of the most important words and phrases that we use in our articles.
We will regularly update this glossary to add new entries as researchers continually uncover novel methods to push the frontier of artificial intelligence while identifying emerging safety risks.
AGI
Artificial general intelligence, or AGI, is a nebulous term. But it generally refers to AI that’s more capable than the average human at many, if not most, tasks. OpenAI CEO Sam Altman
recently
described AGI as the “equivalent of a median human that you could hire as a co-worker.” Meanwhile,
OpenAI’s charter
defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” Google DeepMind’s understanding differs slightly from these two definitions; the lab views AGI as “AI that’s at least as capable as humans at most cognitive tasks.” Confused? Not to worry —
so are experts at the forefront of AI research
.
AI agent
An AI agent refers to a tool that uses AI technologies to perform a series of tasks on your behalf — beyond what a more basic AI chatbot could do — such as filing expenses, booking tickets or a table at a restaurant, or even writing and maintaining code. However, as we’ve
explained before
, there are lots of moving pieces in this emergent space, so “AI agent” might mean different things to different people. Infrastructure is also still being built out to deliver on its envisaged capabilities. But the basic concept implies an autonomous system that may draw on multiple AI systems to carry out multistep tasks.
Chain of thought
Given a simple question, a human brain can answer without even thinking too much about it — things like “which animal is taller, a giraffe or a cat?” But in many cases, you often need a pen and paper to come up with the right answer because there are intermediary steps. For instance, if a farmer has chickens and cows, and together they have 40 heads and 120 legs, you might need to write down a simple equation to come up with the answer (20 chickens and 20 cows).
In an AI context, chain-of-thought reasoning for large language models means breaking down a problem into smaller, intermediate steps to improve the quality of the end result. It usually takes longer to get an answer, but the answer is more likely to be correct, especially in a logic or coding context. Reasoning models are developed from traditional large language models and optimized for chain-of-thought thinking thanks to reinforcement learning.
(See:
Large language model
)
Techcrunch event
San Francisco, CA
|
October 13-15, 2026
Compute
Although somewhat of a multivalent term, compute generally refers to the vital
computational power
that allows AI models to operate. This type of processing fuels the AI industry, giving it the ability to train and deploy its powerful models. The term is often a shorthand for the kinds of hardware that provides the computational power — things like GPUs, CPUs, TPUs, and other forms of infrastructure that form the bedrock of the modern AI industry.
Deep learning
A subset of self-improving machine learning in which AI algorithms are designed with a multi-layered, artificial neural network (ANN) structure. This allows them to make more complex correlations compared to simpler machine learning-based systems, such as linear models or decision trees. The structure of deep learning algorithms draws inspiration from the interconnected pathways of neurons in the human brain.
Deep learning AI models are able to identify important characteristics in data themselves, rather than requiring human engineers to define these features. The structure also supports algorithms that can learn from errors and, through a process of repetition and adjustment, improve their own outputs. However, deep learning systems require a lot of data points to yield good results (millions or more). They also typically take longer to train compared to simpler machine learning algorithms — so development costs tend to be higher.
(See:
Neural network
)
Diffusion
Diffusion is the tech at the heart of many art-, music-, and text-generating AI models. Inspired by physics,
diffusion systems slowly “destroy” the structure of data
— for example, photos, songs, and so on — by adding noise until there’s nothing left. In physics, diffusion is spontaneous and irreversible — sugar diffused in coffee can’t be restored to cube form. But diffusion systems in AI aim to learn a sort of “reverse diffusion” process to restore the destroyed data, gaining the ability to recover the data from noise.
Distillation
Distillation is a technique used to extract knowledge from a large AI model with a ‘teacher-student’ model. Developers send requests to a teacher model and record the outputs. Answers are sometimes compared with a dataset to see how accurate they are. These outputs are then used to train the student model, which is trained to approximate the teacher’s behavior.
Distillation can be used to create a smaller, more efficient model based on a larger model with a minimal distillation loss. This is likely how OpenAI developed GPT-4 Turbo, a faster version of GPT-4.
While all AI companies use distillation internally, it may have also been used by some AI companies to catch up with frontier models. Distillation from a competitor usually
violates
the terms of service of AI API and chat assistants.
Fine-tuning
This refers to the further training of an AI model to optimize performance for a more specific task or area than was previously a focal point of its training — typically by feeding in new, specialized (i.e., task-oriented) data.
Many AI startups are taking large language models as a starting point to build a commercial product but are vying to amp up utility for a target sector or task by supplementing earlier training cycles with fine-tuning based on their own domain-specific knowledge and expertise.
(See:
Large language model [LLM]
)
GAN
A GAN, or Generative Adversarial Network, is a type of machine learning framework that underpins some important developments in generative AI when it comes to producing realistic data — including (but not only) deepfake tools. GANs involve the use of a pair of neural networks, one of which draws on its training data to generate an output that is passed to the other model to evaluate. This second, discriminator model thus plays the role of a classifier on the generator’s output – enabling it to improve over time.
The GAN structure is set up as a competition (hence “adversarial”) – with the two models essentially programmed to try to outdo each other: the generator is trying to get its output past the discriminator, while the discriminator is working to spot artificially generated data. This structured contest can optimize AI outputs to be more realistic without the need for additional human intervention. Though GANs work best for narrower applications (such as producing realistic photos or videos), rather than general purpose AI.
Hallucination
Hallucination is the AI industry’s preferred term for AI models making stuff up – literally generating information that is incorrect. Obviously, it’s a huge problem for AI quality.
Hallucinations produce GenAI outputs that can be misleading and could even lead to real-life risks — with potentially dangerous consequences (think of a health query that returns harmful medical advice). This is why most GenAI tools’ small print now warns users to verify AI-generated answers, even though such disclaimers are usually far less prominent than the information the tools dispense at the touch of a button.
The problem of AIs fabricating information is thought to arise as a consequence of gaps in training data. For general purpose GenAI especially — also sometimes known as foundation models — this looks difficult to resolve. There is simply not enough data in existence to train AI models to comprehensively resolve all the questions we could possibly ask. TL;DR: we haven’t invented God (yet).
Hallucinations are contributing to a push towards increasingly specialized and/or vertical AI models — i.e. domain-specific AIs that require narrower expertise – as a way to reduce the likelihood of knowledge gaps and shrink disinformation risks.
Inference
Inference is the process of running an AI model. It’s setting a model loose to make predictions or draw conclusions from previously seen data. To be clear, inference can’t happen without training; a model must learn patterns in a set of data before it can effectively extrapolate from this training data.
Many types of hardware can perform inference, ranging from smartphone processors to beefy GPUs to custom-designed AI accelerators. But not all of them can run models equally well. Very large models would take ages to make predictions on, say, a laptop versus a cloud server with high-end AI chips.
[See:
Training
]
Large language model (LLM)
Large language models, or LLMs, are the AI models used by popular AI assistants, such as
ChatGPT
,
Claude
,
Google’s Gemini
,
Meta’s AI Llama
,
Microsoft Copilot
, or
Mistral’s Le Chat
. When you chat with an AI assistant, you interact with a large language model that processes your request directly or with the help of different available tools, such as web browsing or code interpreters.
AI assistants and LLMs can have different names. For instance, GPT is OpenAI’s large language model and ChatGPT is the AI assistant product.
LLMs are deep neural networks made of billions of numerical parameters (
or weights, see below
) that learn the relationships between words and phrases and create a representation of language, a sort of multidimensional map of words.
These models are created from encoding the patterns they find in billions of books, articles, and transcripts. When you prompt an LLM, the model generates the most likely pattern that fits the prompt. It then evaluates the most probable next word after the last one based on what was said before. Repeat, repeat, and repeat.
(See:
Neural network
)
Memory cache
Memory cache refers to an important process that boosts inference (which is the process by which AI works to generate a response to a user’s query). In essence, caching is an optimization technique, designed to make inference more efficient. AI is obviously driven by high-octane mathematical calculations and every time those calculations are made, they use up more power. Caching is designed to cut down on the number of calculations a model might have to run by saving particular calculations for future user queries and operations. There are different kinds of memory caching, although one of the more well-known is
KV (or key value) caching
. KV caching works in transformer-based models, and increases efficiency, driving faster results by reducing the amount of time (and algorithmic labor) it takes to generate answers to user questions.
(See:
Inference
)
Neural network
A neural network refers to the multi-layered algorithmic structure that underpins deep learning — and, more broadly, the whole boom in generative AI tools following the emergence of large language models.
Although the idea of taking inspiration from the densely interconnected pathways of the human brain as a design structure for data processing algorithms dates all the way back to the 1940s, it was the much more recent rise of graphical processing hardware (GPUs) — via the video game industry — that really unlocked the power of this theory. These chips proved well suited to training algorithms with many more layers than was possible in earlier epochs — enabling neural network-based AI systems to achieve far better performance across many domains, including voice recognition, autonomous navigation, and drug discovery.
(See:
Large language model [LLM]
)
RAMageddon
RAMageddon is the fun new term for a not-so-fun trend that is sweeping the tech industry: an ever-increasing shortage of random access memory, or RAM chips, which power pretty much all the tech products we use in our daily lives. As the AI industry has blossomed, the biggest tech companies and AI labs — all vying to have the most powerful and efficient AI — are buying so much RAM to power their data centers that there’s not much left for the rest of us. And that supply bottleneck means that what’s left is getting more and more expensive.
That includes industries like gaming (where major companies have had to
raise prices on consoles
because it’s harder to find memory chips for their devices), consumer electronics (where memory shortage could cause
the biggest dip in smartphone shipments
in more than a decade), and general enterprise computing (because those companies can’t get enough RAM for their own data centers). The surge in prices is only expected to stop after the dreaded shortage ends but, unfortunately, there’s
not really much of a sign
that’s going to happen anytime soon.
Training
Developing machine learning AIs involves a process known as training. In simple terms, this refers to data being fed in in order that the model can learn from patterns and generate useful outputs.
Things can get a bit philosophical at this point in the AI stack — since, pre-training, the mathematical structure that’s used as the starting point for developing a learning system is just a bunch of layers and random numbers. It’s only through training that the AI model really takes shape. Essentially, it’s the process of the system responding to characteristics in the data that enables it to adapt outputs towards a sought-for goal — whether that’s identifying images of cats or producing a haiku on demand.
It’s important to note that not all AI requires training. Rules-based AIs that are programmed to follow manually predefined instructions — for example, such as linear chatbots — don’t need to undergo training. However, such AI systems are likely to be more constrained than (well-trained) self-learning systems.
Still, training can be expensive because it requires lots of inputs — and, typically, the volumes of inputs required for such models have been trending upwards.
Hybrid approaches can sometimes be used to shortcut model development and help manage costs. Such as doing data-driven fine-tuning of a rules-based AI — meaning development requires less data, compute, energy, and algorithmic complexity than if the developer had started building from scratch.
[See:
Inference
]
Tokens
When it comes to human-machine communication, there are some obvious challenges. People communicate using human language, while AI programs execute tasks and respond to queries through complex algorithmic processes that are informed by data. In their simplest definition, tokens represent the basic building blocks of human-AI communication, in that they are discrete segments of data that have either been processed or produced by an LLM.
Tokens are created via a process known as “tokenization,” which breaks down raw data and refines it into distinct units that are digestible to an LLM. Similar to how a software compiler translates human language into binary code that a computer can digest, tokenization interprets human language for an AI program via their user queries so that it can prepare a response.
There are several different kinds of tokens — including input tokens (the kind that must be generated in response to a human user’s query), output tokens (the kind that are generated as the LLM responds to the human’s request), and reasoning tokens, which involve longer, more intensive tasks and processes that occur as part of a user request.
With enterprise AI, token usage also determines costs. Since tokens are equivalent to the amount of data being processed by a model, they have also become the means by which the AI industry monetizes its services. Most AI companies charge for LLM usage on a per-token-basis. Thus, the more tokens a business burns as it uses an AI program (ChatGPT, for example), the more money it will have to pay its AI service provider (OpenAI).
Transfer learning
A technique where a previously trained AI model is used as the starting point for developing a new model for a different but typically related task – allowing knowledge gained in previous training cycles to be reapplied.
Transfer learning can drive efficiency savings by shortcutting model development. It can also be useful when data for the task that the model is being developed for is somewhat limited. But it’s important to note that the approach has limitations. Models that rely on transfer learning to gain generalized capabilities will likely require training on additional data in order to perform well in their domain of focus
(See:
Fine tuning
)
Weights
Weights are core to AI training, as they determine how much importance (or weight) is given to different features (or input variables) in the data used for training the system — thereby shaping the AI model’s output.
Put another way, weights are numerical parameters that define what’s most salient in a dataset for the given training task. They achieve their function by applying multiplication to inputs. Model training typically begins with weights that are randomly assigned, but as the process unfolds, the weights adjust as the model seeks to arrive at an output that more closely matches the target.
For example, an AI model for predicting housing prices that’s trained on historical real estate data for a target location could include weights for features such as the number of bedrooms and bathrooms, whether a property is detached or semi-detached, whether it has parking, a garage, and so on.
Ultimately, the weights the model attaches to each of these inputs reflect how much they influence the value of a property, based on the given dataset.
This article is updated regularly with new information. |
| Markdown | [Skip to content](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#wp--skip-link--target)
The first StrictlyVC of 2026 hits San Francisco. Tickets are going fast. [Register now.](https://techcrunch.com/events/strictlyvc-san-francisco-2026/?utm_source=tc&utm_medium=ad&utm_campaign=svcsf2026&utm_content=ticketsales&promo=topbanner&display=)
Close
[ TechCrunch Desktop Logo](https://techcrunch.com/)
[ TechCrunch Mobile Logo](https://techcrunch.com/)
- [Latest](https://techcrunch.com/latest/)
- [Startups](https://techcrunch.com/category/startups/)
- [Venture](https://techcrunch.com/category/venture/)
- [Apple](https://techcrunch.com/tag/apple/)
- [Security](https://techcrunch.com/category/security/)
- [AI](https://techcrunch.com/category/artificial-intelligence/)
- [Apps](https://techcrunch.com/category/apps/)
- [Events](https://techcrunch.com/events/)
- [Podcasts](https://techcrunch.com/podcasts/)
- [Newsletters](https://techcrunch.com/newsletters/)
Site Search Toggle
Mega Menu Toggle
### Topics
[Latest](https://techcrunch.com/latest/)
[AI](https://techcrunch.com/category/artificial-intelligence/)
[Amazon](https://techcrunch.com/tag/amazon/)
[Apps](https://techcrunch.com/category/apps/)
[Biotech & Health](https://techcrunch.com/category/biotech-health/)
[Climate](https://techcrunch.com/category/climate/)
[Cloud Computing](https://techcrunch.com/tag/cloud-computing/)
[Commerce](https://techcrunch.com/category/commerce/)
[Crypto](https://techcrunch.com/category/cryptocurrency/)
[Enterprise](https://techcrunch.com/category/enterprise/)
[EVs](https://techcrunch.com/tag/evs/)
[Fintech](https://techcrunch.com/category/fintech/)
[Fundraising](https://techcrunch.com/category/fundraising/)
[Gadgets](https://techcrunch.com/category/gadgets/)
[Gaming](https://techcrunch.com/category/gaming/)
[Google](https://techcrunch.com/tag/google/)
[Government & Policy](https://techcrunch.com/category/government-policy/)
[Hardware](https://techcrunch.com/category/hardware/)
[Instagram](https://techcrunch.com/tag/instagram/)
[Layoffs](https://techcrunch.com/tag/layoffs/)
[Media & Entertainment](https://techcrunch.com/category/media-entertainment/)
[Meta](https://techcrunch.com/tag/meta/)
[Microsoft](https://techcrunch.com/tag/microsoft/)
[Privacy](https://techcrunch.com/category/privacy/)
[Robotics](https://techcrunch.com/category/robotics/)
[Security](https://techcrunch.com/category/security/)
[Social](https://techcrunch.com/category/social/)
[Space](https://techcrunch.com/category/space/)
[Startups](https://techcrunch.com/category/startups/)
[TikTok](https://techcrunch.com/tag/tiktok/)
[Transportation](https://techcrunch.com/category/transportation/)
[Venture](https://techcrunch.com/category/venture/)
### More from TechCrunch
[Staff](https://techcrunch.com/about-techcrunch/)
[Events](https://techcrunch.com/events/)
[Startup Battlefield](https://techcrunch.com/startup-battlefield/)
[StrictlyVC](https://strictlyvc.com/)
[Newsletters](https://techcrunch.com/newsletters/)
[Podcasts](https://techcrunch.com/podcasts/)
[Videos](https://techcrunch.com/video/)
[Partner Content](https://techcrunch.com/sponsored/)
[TechCrunch Brand Studio](https://techcrunch.com/brand-studio/)
[Crunchboard](https://www.crunchboard.com/)
[Contact Us](https://techcrunch.com/contact-us/)
**Image Credits:**Getty Images
[Transportation](https://techcrunch.com/category/transportation/)
# Black Sesame, Chinese auto chip challenger to Nvidia, burns \$140M a year
[Rita Liao](https://techcrunch.com/author/rita-liao/)
6:30 AM PDT · July 10, 2023
Chinese auto chip maker Black Sesame [has recently filed to go public](https://www1.hkexnews.hk/app/sehk/2023/105498/documents/sehk23063000070.pdf) in Hong Kong, offering a glimpse into the business prospect and challenges of an industry that’s increasingly important amid an autonomous driving boom and China’s stride toward semiconductor independence.
Founded by veterans of Bosch and OmniVision, Black Sesame is seen as one of the domestic players that can potentially replace the likes of Nvidia, Qualcomm and NXP Semiconductors in the auto chip space. The seven-year-old company has seen its revenue grow significantly over the last three years, but its losses have also ballooned. The question, then, is whether Black Sesame can continue to pour money into R\&D until it becomes profitable.
## Nvidia rival
Black Sesame makes system-on-chips (SoCs) to power autonomous driving cars and other intelligent car functionalities. Its most advanced chip, the Huashan A1000 Pro designed for Level 3 autonomous driving (meaning a vehicle can handle all aspects of driving but still requires human intervention if necessary), offers 160+ TOPS, a unit for measuring computing power. It’s also in the progress of developing a version with 200+ TOPS, which will put it on par with Nvidia’s Drive Orin, which [features 254 TOPS](https://nvidianews.nvidia.com/news/nvidia-unveils-drive-thor-centralized-car-computer-unifying-cluster-infotainment-automated-driving-and-parking-in-a-single-cost-saving-system) and has been in production.
While Black Sesame plays a role in helping China achieve independence in auto chips, its own products rely heavily on access to the global supply chain. For instance, it depends on TSMC to manufacture its SoCs. Its production is vulnerable because the U.S. has been pushing to block TSMC from manufacturing for certain Chinese chip design firms, [especially on the higher end](https://www.bloomberg.com/news/articles/2022-10-22/tsmc-said-to-suspend-work-for-chinese-chip-startup-amid-us-curbs?sref=gni836kR).
As stated in the prospectus, Black Sesame’s ability to receive supplies from the fab could be “adversely affected by international trade policies, geopolitics and trade protection measures, including imposition of trade restrictions and sanctions.”
The company’s chips also feature core parts that are dependent on third-party IPs. Though not specified, these IPs could be subject to the U.S.’s expanding semiconductor war on China.
## Increasing losses
Black Sesame’s revenue tripled from 53 million yuan (\$7.33 million) to 165.4 million yuan between 2020 and 2022, but its losses grew to 1 billion yuan (\$140 million) in 2022, a more than 200% increase from 293 million yuan in 2020. It’s not expected to be profitable in the foreseeable future, for it estimates losses this year to “significantly” increase as it’s in the stage of “expanding” its business, which demands substantial R\&D investments. In 2022, its R\&D expense surged to 764.1 million yuan (\$106 million).
Techcrunch event
### Meet your next investor or portfolio startup at Disrupt
#### Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to \$410.
### Meet your next investor or portfolio startup at Disrupt
#### Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to \$410.
San Francisco, CA \| October 13-15, 2026
[REGISTER NOW](https://techcrunch.com/events/tc-disrupt-2026/?utm_source=tc&utm_medium=ad&utm_campaign=disrupt2026&utm_content=tc_inline_eb&promo=tc_inline_eb&display=)
Its gross profit at the end of 2022 was 29.4%, dwarfed by Nvidia’s [enviable 65%](https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-first-quarter-fiscal-2024).
To date, Black Sesame has raised approximately \$115 million from outside investors, including Nio Capital, the venture fund of EV maker Nio’s founder, a state-owned Dongfeng Motor investment vehicle and Bosch’s China-focused fund Boyuan, according to Crunchbase [data](https://www.crunchbase.com/organization/black-sesame-intelligent-technology). As of 2022, the company had total assets worth \$140 million and a runway of 24 months.
Like many companies in China’s critical industries, Black Sesame receives government grants and tax incentives because it operates in the field of automotive SoCs. But it could lose a significant source of funding if these benefits were to end.
Its financial performance also hinges on its major customer, which accounted for as much as 43.5% of its revenue last year. It currently supplies over 30 original equipment manufacturers and Tier 1 suppliers, including FAW Group, Dongfeng, JAC, HYCAN, ECARX, Baidu, Bosch, ZF Group and Marelli.
> [Volkswagen to plough €2.4B into vehicle automation in China and form JV with Horizon Robotics](https://techcrunch.com/2022/10/13/volkswagen-autonomous-driving-china-horizon-robotics-jv/)
Topics
[auto](https://techcrunch.com/tag/auto/), [autonomous vehicles](https://techcrunch.com/tag/autonomous-vehicles/), [black sesame](https://techcrunch.com/tag/black-sesame/), [China](https://techcrunch.com/tag/china/), [China](https://techcrunch.com/region/asia/china/), [chips](https://techcrunch.com/tag/chips/), [nvidia](https://techcrunch.com/tag/nvidia/), [semiconductor](https://techcrunch.com/tag/semiconductor/), [Transportation](https://techcrunch.com/category/transportation/)
Rita Liao
Reporter, China
Rita covered Asia for TechCrunch, with a special interest in Chinese companies going global and web3 projects with real-world applications. Before her previous writing stints with Tech in Asia and TechNode, Rita managed communications for SOSV’s accelerators in Asia. At one point, she worked for a documentary production company and a mindfulness retreat center in New England. She studied political science and visual arts at Bowdoin College. Contact: ritaliao@pm.me
[View Bio](https://techcrunch.com/author/rita-liao/)
April 30
San Francisco, CA
StrictlyVC kicks off the year in SF. Get in the room for unfiltered fireside chats with industry leaders, insider VC insights, and high-value connections that actually move the needle. **Tickets are limited.**
[**REGISTER NOW**](https://techcrunch.com/events/strictlyvc-san-francisco-2026/?utm_source=tc&utm_medium=ad&utm_campaign=svcsf2026&utm_content=ticketsales&promo=rightrail&display=)
## Most Popular
- ### [Sam Altman responds to ‘incendiary’ New Yorker article after attack on his home](https://techcrunch.com/2026/04/11/sam-altman-responds-to-incendiary-new-yorker-article-after-attack-on-his-home/)
- [Anthony Ha](https://techcrunch.com/author/anthony-ha/)
- ### [Anthropic temporarily banned OpenClaw’s creator from accessing Claude](https://techcrunch.com/2026/04/10/anthropic-temporarily-banned-openclaws-creator-from-accessing-claude/)
- [Julie Bort](https://techcrunch.com/author/julie-bort/)
- ### [France to ditch Windows for Linux to reduce reliance on US tech](https://techcrunch.com/2026/04/10/france-to-ditch-windows-for-linux-to-reduce-reliance-on-us-tech/)
- [Zack Whittaker](https://techcrunch.com/author/zack-whittaker/)
- ### [YouTube Premium and YouTube Music are getting more expensive](https://techcrunch.com/2026/04/10/youtube-premium-youtube-music-subscription-price-increase/)
- [Aisha Malik](https://techcrunch.com/author/aisha-malik/)
Techcrunch event
### Meet your next investor or portfolio startup at Disrupt
#### Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to \$410.
### Meet your next investor or portfolio startup at Disrupt
#### Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to \$410.
San Francisco, CA \| October 13-15, 2026
[REGISTER NOW](https://techcrunch.com/events/tc-disrupt-2026/?utm_source=tc&utm_medium=ad&utm_campaign=disrupt2026&utm_content=tc_inline_eb&promo=tc_inline_eb&display=)
- ### [This founder helped build SpaceX’s most powerful rocket engine. Now he’s building a ‘fighter jet for orbit.’](https://techcrunch.com/2026/04/09/portal-raises-50-million-to-build-a-new-kind-of-high-power-rocket-engine/)
- [Tim Fernholz](https://techcrunch.com/author/tim-fernholz/)
- ### [Developer of VeraCrypt encryption software says Windows users may face boot-up issues after Microsoft locked his account](https://techcrunch.com/2026/04/08/veracrypt-encryption-software-windows-microsoft-lock-boot-issues/)
- [Zack Whittaker](https://techcrunch.com/author/zack-whittaker/)
- ### [Apple’s foldable iPhone is on track to launch in September, report says](https://techcrunch.com/2026/04/07/apples-foldable-iphone-is-on-track-to-launch-in-september-report-says/)
- [Aisha Malik](https://techcrunch.com/author/aisha-malik/)
Keep reading
**Image Credits:**Uber
[Transportation](https://techcrunch.com/category/transportation/)
# Uber and Nuro begin testing premium robotaxi service in San Francisco
[Kirsten Korosec](https://techcrunch.com/author/kirsten-korosec/)
2:11 PM PDT · April 13, 2026
If you spot a Lucid Gravity SUV blinged-out with sensors — and a self-driving system developed by Nuro — driving around San Francisco, chances are that’s an Uber employee taking a ride.
Select Uber employees can now request a ride in a Lucid robotaxi through the Uber app, the latest phase of testing ahead of a planned public launch later this year. Nuro, which provided the update [in a blog](https://www.nuro.ai/blog/robotaxi-employee-test-rides-begin) posted Monday, told TechCrunch the vehicles are operating in autonomous mode and have a human safety operator behind the wheel as backup.
While this is far from a public launch, it does signal the companies’ progress since announcing a [partnership and multimillion-dollar investment](https://techcrunch.com/2025/07/17/uber-makes-multi-million-dollar-investment-in-lucid-nuro-to-build-robotaxi-service/) in July 2025. Uber invested \$300 million in Lucid and separately agreed to buy “at least” 20,000 of the EV maker’s new Gravity SUV over the next six years.
Those EVs are equipped with Nuro’s autonomous vehicle system, which is powered by Nvidia’s Drive AGX Thor computer. The Lucid Gravity robotaxi, which was [revealed in January](https://techcrunch.com/2026/01/05/this-is-ubers-new-robotaxi-from-lucid-and-nuro/), is outfitted with high-resolution cameras, solid-state lidar sensors, and radars that help the self-driving system perceive the real-world environment and operate in it.
Uber also invested an undisclosed “multi-hundred-million dollar” amount into Nuro.
The plan is for Uber to own and operate — likely with the help from a third party — the premium robotaxi service. Production of these modified Lucid Gravity vehicles is expected to begin in late 2026, [according to a regulatory filing](https://www.sec.gov/ix?doc=/Archives/edgar/data/0001811210/000110465925068552/tm2520956d2_8k.htm) posted last year.
Nuro completed closed-course testing and started its first public road testing of the autonomous Lucid Gravity SUVs late last year. Nuro now has 100 Lucid Gravity SUVs outfitted with its self-driving system in the engineering fleet, used to gather real-world data and test autonomous driving across multiple U.S. cities and states.
Techcrunch event
### Meet your next investor or portfolio startup at Disrupt
#### Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to \$410.
### Meet your next investor or portfolio startup at Disrupt
#### Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to \$410.
San Francisco, CA \| October 13-15, 2026
[REGISTER NOW](https://techcrunch.com/events/tc-disrupt-2026/?utm_source=tc&utm_medium=ad&utm_campaign=disrupt2026&utm_content=tc_inline_eb&promo=tc_inline_eb&display=)
According to Nuro, the employee test rides help the team evaluate how the autonomy stack, vehicle, and rider experience work together and function in a live operating environment. It also allows the team test how well the vehicle handles rider pickups and drop-offs, a notoriously tricky operation in autonomous ride-hailing.
Topics
[lucid](https://techcrunch.com/tag/lucid/), [nuro](https://techcrunch.com/tag/nuro/), [nvidia](https://techcrunch.com/tag/nvidia/), [robotaxis](https://techcrunch.com/tag/robotaxis/), [Transportation](https://techcrunch.com/category/transportation/), [Uber](https://techcrunch.com/tag/uber/)
Kirsten Korosec
Transportation Editor
Kirsten Korosec is a reporter and editor who has covered the future of transportation from EVs and autonomous vehicles to urban air mobility and in-car tech for more than a decade. She is currently the transportation editor at TechCrunch and co-host of TechCrunch’s Equity podcast. She is also co-founder and co-host of the podcast, “The Autonocast.” She previously wrote for Fortune, The Verge, Bloomberg, MIT Technology Review and CBS Interactive.
You can contact or verify outreach from Kirsten by emailing [kirsten.korosec@techcrunch.com](mailto:kirsten.korosec@techcrunch.com) or via encrypted message at kkorosec.07 on Signal.
[View Bio](https://techcrunch.com/author/kirsten-korosec/)
### Newsletters
[See More](https://techcrunch.com/newsletters/)
Subscribe for the industry’s biggest tech news
Keep reading
**Image Credits:**Getty Images AI Generator / Getty Images
[AI](https://techcrunch.com/category/artificial-intelligence/)
# From LLMs to hallucinations, here’s a simple guide to common AI terms
[Natasha Lomas](https://techcrunch.com/author/natasha-lomas/)
[Romain Dillet](https://techcrunch.com/author/romain-dillet/)
[Kyle Wiggers](https://techcrunch.com/author/kyle-wiggers/)
[Lucas Ropek](https://techcrunch.com/author/lucas-ropek/)
8:07 AM PDT · April 12, 2026
Artificial intelligence is a deep and convoluted world. The scientists who work in this field often rely on jargon and lingo to explain what they’re working on. As a result, we frequently have to use those technical terms in our coverage of the artificial intelligence industry. That’s why we thought it would be helpful to put together a glossary with definitions of some of the most important words and phrases that we use in our articles.
We will regularly update this glossary to add new entries as researchers continually uncover novel methods to push the frontier of artificial intelligence while identifying emerging safety risks.
***
## [AGI](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#agi)
Artificial general intelligence, or AGI, is a nebulous term. But it generally refers to AI that’s more capable than the average human at many, if not most, tasks. OpenAI CEO Sam Altman [recently](https://nymag.com/intelligencer/article/sam-altman-artificial-intelligence-openai-profile.html) described AGI as the “equivalent of a median human that you could hire as a co-worker.” Meanwhile, [OpenAI’s charter](https://openai.com/charter/) defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” Google DeepMind’s understanding differs slightly from these two definitions; the lab views AGI as “AI that’s at least as capable as humans at most cognitive tasks.” Confused? Not to worry — [so are experts at the forefront of AI research](https://techcrunch.com/2024/10/03/even-the-godmother-of-ai-has-no-idea-what-agi-is/).
## [AI agent](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#ai-agent)
An AI agent refers to a tool that uses AI technologies to perform a series of tasks on your behalf — beyond what a more basic AI chatbot could do — such as filing expenses, booking tickets or a table at a restaurant, or even writing and maintaining code. However, as we’ve [explained before](https://techcrunch.com/2024/12/15/what-exactly-is-an-ai-agent/), there are lots of moving pieces in this emergent space, so “AI agent” might mean different things to different people. Infrastructure is also still being built out to deliver on its envisaged capabilities. But the basic concept implies an autonomous system that may draw on multiple AI systems to carry out multistep tasks.
## [Chain of thought](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#chain-of-thought)
Given a simple question, a human brain can answer without even thinking too much about it — things like “which animal is taller, a giraffe or a cat?” But in many cases, you often need a pen and paper to come up with the right answer because there are intermediary steps. For instance, if a farmer has chickens and cows, and together they have 40 heads and 120 legs, you might need to write down a simple equation to come up with the answer (20 chickens and 20 cows).
In an AI context, chain-of-thought reasoning for large language models means breaking down a problem into smaller, intermediate steps to improve the quality of the end result. It usually takes longer to get an answer, but the answer is more likely to be correct, especially in a logic or coding context. Reasoning models are developed from traditional large language models and optimized for chain-of-thought thinking thanks to reinforcement learning.
(See: [Large language model](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#large-language-model))
Techcrunch event
### Meet your next investor or portfolio startup at Disrupt
#### Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to \$410.
### Meet your next investor or portfolio startup at Disrupt
#### Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to \$410.
San Francisco, CA \| October 13-15, 2026
[REGISTER NOW](https://techcrunch.com/events/tc-disrupt-2026/?utm_source=tc&utm_medium=ad&utm_campaign=disrupt2026&utm_content=tc_inline_eb&promo=tc_inline_eb&display=)
## [Compute](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#compute)
Although somewhat of a multivalent term, compute generally refers to the vital [computational power](https://carnegieendowment.org/posts/2024/04/a-primer-on-compute) that allows AI models to operate. This type of processing fuels the AI industry, giving it the ability to train and deploy its powerful models. The term is often a shorthand for the kinds of hardware that provides the computational power — things like GPUs, CPUs, TPUs, and other forms of infrastructure that form the bedrock of the modern AI industry.
## [Deep learning](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#deep-learning)
A subset of self-improving machine learning in which AI algorithms are designed with a multi-layered, artificial neural network (ANN) structure. This allows them to make more complex correlations compared to simpler machine learning-based systems, such as linear models or decision trees. The structure of deep learning algorithms draws inspiration from the interconnected pathways of neurons in the human brain.
Deep learning AI models are able to identify important characteristics in data themselves, rather than requiring human engineers to define these features. The structure also supports algorithms that can learn from errors and, through a process of repetition and adjustment, improve their own outputs. However, deep learning systems require a lot of data points to yield good results (millions or more). They also typically take longer to train compared to simpler machine learning algorithms — so development costs tend to be higher.
(See: [Neural network](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#neural-network))
## [Diffusion](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#diffusion)
Diffusion is the tech at the heart of many art-, music-, and text-generating AI models. Inspired by physics, [diffusion systems slowly “destroy” the structure of data](https://techcrunch.com/2022/12/22/a-brief-history-of-diffusion-the-tech-at-the-heart-of-modern-image-generating-ai/) — for example, photos, songs, and so on — by adding noise until there’s nothing left. In physics, diffusion is spontaneous and irreversible — sugar diffused in coffee can’t be restored to cube form. But diffusion systems in AI aim to learn a sort of “reverse diffusion” process to restore the destroyed data, gaining the ability to recover the data from noise.
## [Distillation](http://distillation/)
Distillation is a technique used to extract knowledge from a large AI model with a ‘teacher-student’ model. Developers send requests to a teacher model and record the outputs. Answers are sometimes compared with a dataset to see how accurate they are. These outputs are then used to train the student model, which is trained to approximate the teacher’s behavior.
Distillation can be used to create a smaller, more efficient model based on a larger model with a minimal distillation loss. This is likely how OpenAI developed GPT-4 Turbo, a faster version of GPT-4.
While all AI companies use distillation internally, it may have also been used by some AI companies to catch up with frontier models. Distillation from a competitor usually [violates](https://techcrunch.com/2025/01/29/microsoft-probing-whether-deepseek-improperly-used-openais-api/) the terms of service of AI API and chat assistants.
## [Fine-tuning](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#fine-tuning)
This refers to the further training of an AI model to optimize performance for a more specific task or area than was previously a focal point of its training — typically by feeding in new, specialized (i.e., task-oriented) data.
Many AI startups are taking large language models as a starting point to build a commercial product but are vying to amp up utility for a target sector or task by supplementing earlier training cycles with fine-tuning based on their own domain-specific knowledge and expertise.
(See: [Large language model \[LLM\]](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#large-language-model))
## [GAN](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#gan)
A GAN, or Generative Adversarial Network, is a type of machine learning framework that underpins some important developments in generative AI when it comes to producing realistic data — including (but not only) deepfake tools. GANs involve the use of a pair of neural networks, one of which draws on its training data to generate an output that is passed to the other model to evaluate. This second, discriminator model thus plays the role of a classifier on the generator’s output – enabling it to improve over time.
The GAN structure is set up as a competition (hence “adversarial”) – with the two models essentially programmed to try to outdo each other: the generator is trying to get its output past the discriminator, while the discriminator is working to spot artificially generated data. This structured contest can optimize AI outputs to be more realistic without the need for additional human intervention. Though GANs work best for narrower applications (such as producing realistic photos or videos), rather than general purpose AI.
## [Hallucination](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#hallucination)
Hallucination is the AI industry’s preferred term for AI models making stuff up – literally generating information that is incorrect. Obviously, it’s a huge problem for AI quality.
Hallucinations produce GenAI outputs that can be misleading and could even lead to real-life risks — with potentially dangerous consequences (think of a health query that returns harmful medical advice). This is why most GenAI tools’ small print now warns users to verify AI-generated answers, even though such disclaimers are usually far less prominent than the information the tools dispense at the touch of a button.
The problem of AIs fabricating information is thought to arise as a consequence of gaps in training data. For general purpose GenAI especially — also sometimes known as foundation models — this looks difficult to resolve. There is simply not enough data in existence to train AI models to comprehensively resolve all the questions we could possibly ask. TL;DR: we haven’t invented God (yet).
Hallucinations are contributing to a push towards increasingly specialized and/or vertical AI models — i.e. domain-specific AIs that require narrower expertise – as a way to reduce the likelihood of knowledge gaps and shrink disinformation risks.
## [Inference](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#inference)
Inference is the process of running an AI model. It’s setting a model loose to make predictions or draw conclusions from previously seen data. To be clear, inference can’t happen without training; a model must learn patterns in a set of data before it can effectively extrapolate from this training data.
Many types of hardware can perform inference, ranging from smartphone processors to beefy GPUs to custom-designed AI accelerators. But not all of them can run models equally well. Very large models would take ages to make predictions on, say, a laptop versus a cloud server with high-end AI chips.
\[See: [Training](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#training)\]
## [Large language model (LLM)](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#large-language-model)
Large language models, or LLMs, are the AI models used by popular AI assistants, such as [ChatGPT](https://techcrunch.com/2025/02/12/chatgpt-everything-to-know-about-the-ai-chatbot/), [Claude](https://techcrunch.com/2025/02/25/claude-everything-you-need-to-know-about-anthropics-ai/), [Google’s Gemini](https://techcrunch.com/2025/02/26/what-is-google-gemini-ai/), [Meta’s AI Llama](https://techcrunch.com/2024/09/08/meta-llama-everything-you-need-to-know-about-the-open-generative-ai-model/), [Microsoft Copilot](https://techcrunch.com/2024/08/17/microsoft-copilot-everything-you-need-to-know-about-microsofts-ai/), or [Mistral’s Le Chat](https://techcrunch.com/2025/02/28/what-is-mistral-ai-everything-to-know-about-the-openai-competitor/). When you chat with an AI assistant, you interact with a large language model that processes your request directly or with the help of different available tools, such as web browsing or code interpreters.
AI assistants and LLMs can have different names. For instance, GPT is OpenAI’s large language model and ChatGPT is the AI assistant product.
LLMs are deep neural networks made of billions of numerical parameters ([or weights, see below](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#weights)) that learn the relationships between words and phrases and create a representation of language, a sort of multidimensional map of words.
These models are created from encoding the patterns they find in billions of books, articles, and transcripts. When you prompt an LLM, the model generates the most likely pattern that fits the prompt. It then evaluates the most probable next word after the last one based on what was said before. Repeat, repeat, and repeat.
(See: [Neural network](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#neural-network))
## [Memory cache](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#memory-cache)
Memory cache refers to an important process that boosts inference (which is the process by which AI works to generate a response to a user’s query). In essence, caching is an optimization technique, designed to make inference more efficient. AI is obviously driven by high-octane mathematical calculations and every time those calculations are made, they use up more power. Caching is designed to cut down on the number of calculations a model might have to run by saving particular calculations for future user queries and operations. There are different kinds of memory caching, although one of the more well-known is [KV (or key value) caching](https://huggingface.co/blog/not-lain/kv-caching#:~:text=References%20&%20Further%20Reading-,Introduction,much%20faster%20and%20more%20efficient.). KV caching works in transformer-based models, and increases efficiency, driving faster results by reducing the amount of time (and algorithmic labor) it takes to generate answers to user questions.
(See: [Inference](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#inference))
## [Neural network](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#neural-network)
A neural network refers to the multi-layered algorithmic structure that underpins deep learning — and, more broadly, the whole boom in generative AI tools following the emergence of large language models.
Although the idea of taking inspiration from the densely interconnected pathways of the human brain as a design structure for data processing algorithms dates all the way back to the 1940s, it was the much more recent rise of graphical processing hardware (GPUs) — via the video game industry — that really unlocked the power of this theory. These chips proved well suited to training algorithms with many more layers than was possible in earlier epochs — enabling neural network-based AI systems to achieve far better performance across many domains, including voice recognition, autonomous navigation, and drug discovery.
(See: [Large language model \[LLM\]](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#large-language-model))
## [RAMageddon](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#ramageddon)
RAMageddon is the fun new term for a not-so-fun trend that is sweeping the tech industry: an ever-increasing shortage of random access memory, or RAM chips, which power pretty much all the tech products we use in our daily lives. As the AI industry has blossomed, the biggest tech companies and AI labs — all vying to have the most powerful and efficient AI — are buying so much RAM to power their data centers that there’s not much left for the rest of us. And that supply bottleneck means that what’s left is getting more and more expensive.
That includes industries like gaming (where major companies have had to [raise prices on consoles](https://gizmodo.com/game-consoles-will-likely-get-even-more-expensive-again-2000722792) because it’s harder to find memory chips for their devices), consumer electronics (where memory shortage could cause [the biggest dip in smartphone shipments](https://techcrunch.com/2026/02/27/memory-shortage-could-cause-the-biggest-smartphone-shipments-dip-in-over-a-decade/) in more than a decade), and general enterprise computing (because those companies can’t get enough RAM for their own data centers). The surge in prices is only expected to stop after the dreaded shortage ends but, unfortunately, there’s [not really much of a sign](https://www.tomsguide.com/computing/ram-price-crisis-2026-everything-you-need-to-know) that’s going to happen anytime soon.
## [Training](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#training)
Developing machine learning AIs involves a process known as training. In simple terms, this refers to data being fed in in order that the model can learn from patterns and generate useful outputs.
Things can get a bit philosophical at this point in the AI stack — since, pre-training, the mathematical structure that’s used as the starting point for developing a learning system is just a bunch of layers and random numbers. It’s only through training that the AI model really takes shape. Essentially, it’s the process of the system responding to characteristics in the data that enables it to adapt outputs towards a sought-for goal — whether that’s identifying images of cats or producing a haiku on demand.
It’s important to note that not all AI requires training. Rules-based AIs that are programmed to follow manually predefined instructions — for example, such as linear chatbots — don’t need to undergo training. However, such AI systems are likely to be more constrained than (well-trained) self-learning systems.
Still, training can be expensive because it requires lots of inputs — and, typically, the volumes of inputs required for such models have been trending upwards.
Hybrid approaches can sometimes be used to shortcut model development and help manage costs. Such as doing data-driven fine-tuning of a rules-based AI — meaning development requires less data, compute, energy, and algorithmic complexity than if the developer had started building from scratch.
\[See: [Inference](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#inference)\]
## [Tokens](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#tokens)
When it comes to human-machine communication, there are some obvious challenges. People communicate using human language, while AI programs execute tasks and respond to queries through complex algorithmic processes that are informed by data. In their simplest definition, tokens represent the basic building blocks of human-AI communication, in that they are discrete segments of data that have either been processed or produced by an LLM.
Tokens are created via a process known as “tokenization,” which breaks down raw data and refines it into distinct units that are digestible to an LLM. Similar to how a software compiler translates human language into binary code that a computer can digest, tokenization interprets human language for an AI program via their user queries so that it can prepare a response.
There are several different kinds of tokens — including input tokens (the kind that must be generated in response to a human user’s query), output tokens (the kind that are generated as the LLM responds to the human’s request), and reasoning tokens, which involve longer, more intensive tasks and processes that occur as part of a user request.
With enterprise AI, token usage also determines costs. Since tokens are equivalent to the amount of data being processed by a model, they have also become the means by which the AI industry monetizes its services. Most AI companies charge for LLM usage on a per-token-basis. Thus, the more tokens a business burns as it uses an AI program (ChatGPT, for example), the more money it will have to pay its AI service provider (OpenAI).
## [Transfer learning](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#transfer-learning)
A technique where a previously trained AI model is used as the starting point for developing a new model for a different but typically related task – allowing knowledge gained in previous training cycles to be reapplied.
Transfer learning can drive efficiency savings by shortcutting model development. It can also be useful when data for the task that the model is being developed for is somewhat limited. But it’s important to note that the approach has limitations. Models that rely on transfer learning to gain generalized capabilities will likely require training on additional data in order to perform well in their domain of focus
(See: [Fine tuning](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#fine-tuning))
## [Weights](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#weights)
Weights are core to AI training, as they determine how much importance (or weight) is given to different features (or input variables) in the data used for training the system — thereby shaping the AI model’s output.
Put another way, weights are numerical parameters that define what’s most salient in a dataset for the given training task. They achieve their function by applying multiplication to inputs. Model training typically begins with weights that are randomly assigned, but as the process unfolds, the weights adjust as the model seeks to arrive at an output that more closely matches the target.
For example, an AI model for predicting housing prices that’s trained on historical real estate data for a target location could include weights for features such as the number of bedrooms and bathrooms, whether a property is detached or semi-detached, whether it has parking, a garage, and so on.
Ultimately, the weights the model attaches to each of these inputs reflect how much they influence the value of a property, based on the given dataset.
*This article is updated regularly with new information.*
Topics
[AI](https://techcrunch.com/category/artificial-intelligence/), [artificial intelligence](https://techcrunch.com/tag/artificial-intelligence-2/), [evergreens](https://techcrunch.com/tag/evergreens/), [Glossary](https://techcrunch.com/tag/glossary/)
Natasha Lomas
Senior Reporter
Natasha was a senior reporter for TechCrunch, from September 2012 to April 2025, based in Europe. She joined TC after a stint reviewing smartphones for CNET UK and, prior to that, more than five years covering business technology for silicon.com (now folded into TechRepublic), where she focused on mobile and wireless, telecoms & networking, and IT skills issues. She has also freelanced for organisations including The Guardian and the BBC. Natasha holds a First Class degree in English from Cambridge University, and an MA in journalism from Goldsmiths College, University of London.
[View Bio](https://techcrunch.com/author/natasha-lomas/)
Romain Dillet
Senior Reporter
Romain Dillet was a Senior Reporter at TechCrunch until April 2025. He has written over 3,500 articles on technology and tech startups and has established himself as an influential voice on the European tech scene. He has a deep background in startups, AI, fintech, privacy, security, blockchain, mobile, social and media. With thirteen years of experience at TechCrunch, he’s one of the familiar faces of the tech publication that obsessively covers Silicon Valley and the tech industry — his career started at TechCrunch when he was 21. Based in Paris, many people in the tech ecosystem consider him as the most knowledgeable tech journalist in town. Romain likes to spot important startups before anyone else. He was the first person to cover Revolut, Alan and N26. He has written scoops on large acquisitions from Apple, Microsoft and Snap. When he’s not writing, Romain is also a developer — he understands how the tech behind the tech works. He also has a deep historical knowledge of the computer industry for the past 50 years. He knows how to connect the dots between innovations and the effect on the fabric of our society. Romain graduated from Emlyon Business School, a leading French business school specialized in entrepreneurship. He has helped several non-profit organizations, such as StartHer, an organization that promotes education and empowerment of women in technology, and Techfugees, an organization that empowers displaced people with technology.
[View Bio](https://techcrunch.com/author/romain-dillet/)
Kyle Wiggers
AI Editor
Kyle Wiggers was TechCrunch’s AI Editor until June 2025. His writing has appeared in VentureBeat and Digital Trends, as well as a range of gadget blogs including Android Police, Android Authority, Droid-Life, and XDA-Developers. He lives in Manhattan with his partner, a music therapist.
[View Bio](https://techcrunch.com/author/kyle-wiggers/)
Lucas Ropek
Senior Writer, TechCrunch
Lucas is a senior writer at TechCrunch, where he covers artificial intelligence, consumer tech, and startups. He previously covered AI and cybersecurity at Gizmodo. You can contact Lucas by emailing lucas.ropek@techcrunch.com.
[View Bio](https://techcrunch.com/author/lucas-ropek/)
## Latest in AI
[See More](https://techcrunch.com/category/artificial-intelligence/)
- 
[Startups](https://techcrunch.com/category/startups/)
### [OpenAI has bought AI personal finance startup Hiro](https://techcrunch.com/2026/04/13/openai-has-bought-ai-personal-finance-startup-hiro/)
- [Julie Bort](https://techcrunch.com/author/julie-bort/)
2 hours ago
- 
[Enterprise](https://techcrunch.com/category/enterprise/)
### [Microsoft is working on yet another OpenClaw-like agent](https://techcrunch.com/2026/04/13/microsoft-is-working-on-yet-another-openclaw-like-agent/)
- [Julie Bort](https://techcrunch.com/author/julie-bort/)
7 hours ago
- 
[AI](https://techcrunch.com/category/artificial-intelligence/)
### [Stanford report highlights growing disconnect between AI insiders and everyone else](https://techcrunch.com/2026/04/13/stanford-report-highlights-growing-disconnect-between-ai-insiders-and-everyone-else/)
- [Sarah Perez](https://techcrunch.com/author/sarah-perez/)
8 hours ago
Loading the next article
Error loading the next article
✕
[](https://techcrunch.com/)
- [X](https://twitter.com/techcrunch)
- [LinkedIn](https://www.linkedin.com/company/techcrunch)
- [Facebook](https://www.facebook.com/techcrunch)
- [Instagram](https://instagram.com/techcrunch)
- [youTube](https://www.youtube.com/user/techcrunch)
- [Mastodon](https://mstdn.social/@TechCrunch)
- [Threads](https://www.threads.net/@techcrunch)
- [Bluesky](https://bsky.app/profile/techcrunch.com)
- [TechCrunch](https://techcrunch.com/)
- [Staff](https://techcrunch.com/about-techcrunch/)
- [Contact Us](https://techcrunch.com/contact-us/)
- [Advertise](https://techcrunch.com/advertise/)
- [Crunchboard Jobs](https://www.crunchboard.com/)
- [Site Map](https://techcrunch.com/site-map/)
- [Terms of Service](https://techcrunch.com/terms-of-service/)
- [Privacy Policy](https://techcrunch.com/privacy-policy/)
- [RSS Terms of Use](https://techcrunch.com/rss-terms-of-use/)
- [Code of Conduct](https://techcrunch.com/code-of-conduct/)
- [OpenAI](https://techcrunch.com/2026/04/06/openais-vision-for-the-ai-economy-public-wealth-funds-robot-taxes-and-a-four-day-work-week/)
- [Iran](https://techcrunch.com/2026/04/06/iran-threatens-stargate-ai-data-centers/)
- [Gas Prices](https://techcrunch.com/2026/04/06/gas-prices-arent-the-only-factor-fueling-used-ev-sales/)
- [Tesla](https://techcrunch.com/2026/04/06/tesla-actually-smart-summon-nhtsa-investigation-smart-parking/)
- [Apple](https://techcrunch.com/2026/04/06/apple-epic-games-lawsuit-supreme-court-appeal-app-store-commission/)
- [Tech Layoffs](https://techcrunch.com/2025/02/28/tech-layoffs-2024-list/)
- [ChatGPT](https://techcrunch.com/2025/01/28/chatgpt-everything-to-know-about-the-ai-chatbot/)
© 2026 TechCrunch Media LLC.
Some areas of this page may shift around if you resize the browser window. Be sure to check heading and document order. |
| Readable Markdown | Artificial intelligence is a deep and convoluted world. The scientists who work in this field often rely on jargon and lingo to explain what they’re working on. As a result, we frequently have to use those technical terms in our coverage of the artificial intelligence industry. That’s why we thought it would be helpful to put together a glossary with definitions of some of the most important words and phrases that we use in our articles.
We will regularly update this glossary to add new entries as researchers continually uncover novel methods to push the frontier of artificial intelligence while identifying emerging safety risks.
***
## [AGI](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#agi)
Artificial general intelligence, or AGI, is a nebulous term. But it generally refers to AI that’s more capable than the average human at many, if not most, tasks. OpenAI CEO Sam Altman [recently](https://nymag.com/intelligencer/article/sam-altman-artificial-intelligence-openai-profile.html) described AGI as the “equivalent of a median human that you could hire as a co-worker.” Meanwhile, [OpenAI’s charter](https://openai.com/charter/) defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” Google DeepMind’s understanding differs slightly from these two definitions; the lab views AGI as “AI that’s at least as capable as humans at most cognitive tasks.” Confused? Not to worry — [so are experts at the forefront of AI research](https://techcrunch.com/2024/10/03/even-the-godmother-of-ai-has-no-idea-what-agi-is/).
## [AI agent](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#ai-agent)
An AI agent refers to a tool that uses AI technologies to perform a series of tasks on your behalf — beyond what a more basic AI chatbot could do — such as filing expenses, booking tickets or a table at a restaurant, or even writing and maintaining code. However, as we’ve [explained before](https://techcrunch.com/2024/12/15/what-exactly-is-an-ai-agent/), there are lots of moving pieces in this emergent space, so “AI agent” might mean different things to different people. Infrastructure is also still being built out to deliver on its envisaged capabilities. But the basic concept implies an autonomous system that may draw on multiple AI systems to carry out multistep tasks.
## [Chain of thought](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#chain-of-thought)
Given a simple question, a human brain can answer without even thinking too much about it — things like “which animal is taller, a giraffe or a cat?” But in many cases, you often need a pen and paper to come up with the right answer because there are intermediary steps. For instance, if a farmer has chickens and cows, and together they have 40 heads and 120 legs, you might need to write down a simple equation to come up with the answer (20 chickens and 20 cows).
In an AI context, chain-of-thought reasoning for large language models means breaking down a problem into smaller, intermediate steps to improve the quality of the end result. It usually takes longer to get an answer, but the answer is more likely to be correct, especially in a logic or coding context. Reasoning models are developed from traditional large language models and optimized for chain-of-thought thinking thanks to reinforcement learning.
(See: [Large language model](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#large-language-model))
Techcrunch event
San Francisco, CA \| October 13-15, 2026
## [Compute](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#compute)
Although somewhat of a multivalent term, compute generally refers to the vital [computational power](https://carnegieendowment.org/posts/2024/04/a-primer-on-compute) that allows AI models to operate. This type of processing fuels the AI industry, giving it the ability to train and deploy its powerful models. The term is often a shorthand for the kinds of hardware that provides the computational power — things like GPUs, CPUs, TPUs, and other forms of infrastructure that form the bedrock of the modern AI industry.
## [Deep learning](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#deep-learning)
A subset of self-improving machine learning in which AI algorithms are designed with a multi-layered, artificial neural network (ANN) structure. This allows them to make more complex correlations compared to simpler machine learning-based systems, such as linear models or decision trees. The structure of deep learning algorithms draws inspiration from the interconnected pathways of neurons in the human brain.
Deep learning AI models are able to identify important characteristics in data themselves, rather than requiring human engineers to define these features. The structure also supports algorithms that can learn from errors and, through a process of repetition and adjustment, improve their own outputs. However, deep learning systems require a lot of data points to yield good results (millions or more). They also typically take longer to train compared to simpler machine learning algorithms — so development costs tend to be higher.
(See: [Neural network](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#neural-network))
## [Diffusion](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#diffusion)
Diffusion is the tech at the heart of many art-, music-, and text-generating AI models. Inspired by physics, [diffusion systems slowly “destroy” the structure of data](https://techcrunch.com/2022/12/22/a-brief-history-of-diffusion-the-tech-at-the-heart-of-modern-image-generating-ai/) — for example, photos, songs, and so on — by adding noise until there’s nothing left. In physics, diffusion is spontaneous and irreversible — sugar diffused in coffee can’t be restored to cube form. But diffusion systems in AI aim to learn a sort of “reverse diffusion” process to restore the destroyed data, gaining the ability to recover the data from noise.
## [Distillation](http://distillation/)
Distillation is a technique used to extract knowledge from a large AI model with a ‘teacher-student’ model. Developers send requests to a teacher model and record the outputs. Answers are sometimes compared with a dataset to see how accurate they are. These outputs are then used to train the student model, which is trained to approximate the teacher’s behavior.
Distillation can be used to create a smaller, more efficient model based on a larger model with a minimal distillation loss. This is likely how OpenAI developed GPT-4 Turbo, a faster version of GPT-4.
While all AI companies use distillation internally, it may have also been used by some AI companies to catch up with frontier models. Distillation from a competitor usually [violates](https://techcrunch.com/2025/01/29/microsoft-probing-whether-deepseek-improperly-used-openais-api/) the terms of service of AI API and chat assistants.
## [Fine-tuning](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#fine-tuning)
This refers to the further training of an AI model to optimize performance for a more specific task or area than was previously a focal point of its training — typically by feeding in new, specialized (i.e., task-oriented) data.
Many AI startups are taking large language models as a starting point to build a commercial product but are vying to amp up utility for a target sector or task by supplementing earlier training cycles with fine-tuning based on their own domain-specific knowledge and expertise.
(See: [Large language model \[LLM\]](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#large-language-model))
## [GAN](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#gan)
A GAN, or Generative Adversarial Network, is a type of machine learning framework that underpins some important developments in generative AI when it comes to producing realistic data — including (but not only) deepfake tools. GANs involve the use of a pair of neural networks, one of which draws on its training data to generate an output that is passed to the other model to evaluate. This second, discriminator model thus plays the role of a classifier on the generator’s output – enabling it to improve over time.
The GAN structure is set up as a competition (hence “adversarial”) – with the two models essentially programmed to try to outdo each other: the generator is trying to get its output past the discriminator, while the discriminator is working to spot artificially generated data. This structured contest can optimize AI outputs to be more realistic without the need for additional human intervention. Though GANs work best for narrower applications (such as producing realistic photos or videos), rather than general purpose AI.
## [Hallucination](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#hallucination)
Hallucination is the AI industry’s preferred term for AI models making stuff up – literally generating information that is incorrect. Obviously, it’s a huge problem for AI quality.
Hallucinations produce GenAI outputs that can be misleading and could even lead to real-life risks — with potentially dangerous consequences (think of a health query that returns harmful medical advice). This is why most GenAI tools’ small print now warns users to verify AI-generated answers, even though such disclaimers are usually far less prominent than the information the tools dispense at the touch of a button.
The problem of AIs fabricating information is thought to arise as a consequence of gaps in training data. For general purpose GenAI especially — also sometimes known as foundation models — this looks difficult to resolve. There is simply not enough data in existence to train AI models to comprehensively resolve all the questions we could possibly ask. TL;DR: we haven’t invented God (yet).
Hallucinations are contributing to a push towards increasingly specialized and/or vertical AI models — i.e. domain-specific AIs that require narrower expertise – as a way to reduce the likelihood of knowledge gaps and shrink disinformation risks.
## [Inference](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#inference)
Inference is the process of running an AI model. It’s setting a model loose to make predictions or draw conclusions from previously seen data. To be clear, inference can’t happen without training; a model must learn patterns in a set of data before it can effectively extrapolate from this training data.
Many types of hardware can perform inference, ranging from smartphone processors to beefy GPUs to custom-designed AI accelerators. But not all of them can run models equally well. Very large models would take ages to make predictions on, say, a laptop versus a cloud server with high-end AI chips.
\[See: [Training](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#training)\]
## [Large language model (LLM)](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#large-language-model)
Large language models, or LLMs, are the AI models used by popular AI assistants, such as [ChatGPT](https://techcrunch.com/2025/02/12/chatgpt-everything-to-know-about-the-ai-chatbot/), [Claude](https://techcrunch.com/2025/02/25/claude-everything-you-need-to-know-about-anthropics-ai/), [Google’s Gemini](https://techcrunch.com/2025/02/26/what-is-google-gemini-ai/), [Meta’s AI Llama](https://techcrunch.com/2024/09/08/meta-llama-everything-you-need-to-know-about-the-open-generative-ai-model/), [Microsoft Copilot](https://techcrunch.com/2024/08/17/microsoft-copilot-everything-you-need-to-know-about-microsofts-ai/), or [Mistral’s Le Chat](https://techcrunch.com/2025/02/28/what-is-mistral-ai-everything-to-know-about-the-openai-competitor/). When you chat with an AI assistant, you interact with a large language model that processes your request directly or with the help of different available tools, such as web browsing or code interpreters.
AI assistants and LLMs can have different names. For instance, GPT is OpenAI’s large language model and ChatGPT is the AI assistant product.
LLMs are deep neural networks made of billions of numerical parameters ([or weights, see below](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#weights)) that learn the relationships between words and phrases and create a representation of language, a sort of multidimensional map of words.
These models are created from encoding the patterns they find in billions of books, articles, and transcripts. When you prompt an LLM, the model generates the most likely pattern that fits the prompt. It then evaluates the most probable next word after the last one based on what was said before. Repeat, repeat, and repeat.
(See: [Neural network](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#neural-network))
## [Memory cache](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#memory-cache)
Memory cache refers to an important process that boosts inference (which is the process by which AI works to generate a response to a user’s query). In essence, caching is an optimization technique, designed to make inference more efficient. AI is obviously driven by high-octane mathematical calculations and every time those calculations are made, they use up more power. Caching is designed to cut down on the number of calculations a model might have to run by saving particular calculations for future user queries and operations. There are different kinds of memory caching, although one of the more well-known is [KV (or key value) caching](https://huggingface.co/blog/not-lain/kv-caching#:~:text=References%20&%20Further%20Reading-,Introduction,much%20faster%20and%20more%20efficient.). KV caching works in transformer-based models, and increases efficiency, driving faster results by reducing the amount of time (and algorithmic labor) it takes to generate answers to user questions.
(See: [Inference](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#inference))
## [Neural network](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#neural-network)
A neural network refers to the multi-layered algorithmic structure that underpins deep learning — and, more broadly, the whole boom in generative AI tools following the emergence of large language models.
Although the idea of taking inspiration from the densely interconnected pathways of the human brain as a design structure for data processing algorithms dates all the way back to the 1940s, it was the much more recent rise of graphical processing hardware (GPUs) — via the video game industry — that really unlocked the power of this theory. These chips proved well suited to training algorithms with many more layers than was possible in earlier epochs — enabling neural network-based AI systems to achieve far better performance across many domains, including voice recognition, autonomous navigation, and drug discovery.
(See: [Large language model \[LLM\]](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#large-language-model))
## [RAMageddon](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#ramageddon)
RAMageddon is the fun new term for a not-so-fun trend that is sweeping the tech industry: an ever-increasing shortage of random access memory, or RAM chips, which power pretty much all the tech products we use in our daily lives. As the AI industry has blossomed, the biggest tech companies and AI labs — all vying to have the most powerful and efficient AI — are buying so much RAM to power their data centers that there’s not much left for the rest of us. And that supply bottleneck means that what’s left is getting more and more expensive.
That includes industries like gaming (where major companies have had to [raise prices on consoles](https://gizmodo.com/game-consoles-will-likely-get-even-more-expensive-again-2000722792) because it’s harder to find memory chips for their devices), consumer electronics (where memory shortage could cause [the biggest dip in smartphone shipments](https://techcrunch.com/2026/02/27/memory-shortage-could-cause-the-biggest-smartphone-shipments-dip-in-over-a-decade/) in more than a decade), and general enterprise computing (because those companies can’t get enough RAM for their own data centers). The surge in prices is only expected to stop after the dreaded shortage ends but, unfortunately, there’s [not really much of a sign](https://www.tomsguide.com/computing/ram-price-crisis-2026-everything-you-need-to-know) that’s going to happen anytime soon.
## [Training](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#training)
Developing machine learning AIs involves a process known as training. In simple terms, this refers to data being fed in in order that the model can learn from patterns and generate useful outputs.
Things can get a bit philosophical at this point in the AI stack — since, pre-training, the mathematical structure that’s used as the starting point for developing a learning system is just a bunch of layers and random numbers. It’s only through training that the AI model really takes shape. Essentially, it’s the process of the system responding to characteristics in the data that enables it to adapt outputs towards a sought-for goal — whether that’s identifying images of cats or producing a haiku on demand.
It’s important to note that not all AI requires training. Rules-based AIs that are programmed to follow manually predefined instructions — for example, such as linear chatbots — don’t need to undergo training. However, such AI systems are likely to be more constrained than (well-trained) self-learning systems.
Still, training can be expensive because it requires lots of inputs — and, typically, the volumes of inputs required for such models have been trending upwards.
Hybrid approaches can sometimes be used to shortcut model development and help manage costs. Such as doing data-driven fine-tuning of a rules-based AI — meaning development requires less data, compute, energy, and algorithmic complexity than if the developer had started building from scratch.
\[See: [Inference](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#inference)\]
## [Tokens](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#tokens)
When it comes to human-machine communication, there are some obvious challenges. People communicate using human language, while AI programs execute tasks and respond to queries through complex algorithmic processes that are informed by data. In their simplest definition, tokens represent the basic building blocks of human-AI communication, in that they are discrete segments of data that have either been processed or produced by an LLM.
Tokens are created via a process known as “tokenization,” which breaks down raw data and refines it into distinct units that are digestible to an LLM. Similar to how a software compiler translates human language into binary code that a computer can digest, tokenization interprets human language for an AI program via their user queries so that it can prepare a response.
There are several different kinds of tokens — including input tokens (the kind that must be generated in response to a human user’s query), output tokens (the kind that are generated as the LLM responds to the human’s request), and reasoning tokens, which involve longer, more intensive tasks and processes that occur as part of a user request.
With enterprise AI, token usage also determines costs. Since tokens are equivalent to the amount of data being processed by a model, they have also become the means by which the AI industry monetizes its services. Most AI companies charge for LLM usage on a per-token-basis. Thus, the more tokens a business burns as it uses an AI program (ChatGPT, for example), the more money it will have to pay its AI service provider (OpenAI).
## [Transfer learning](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#transfer-learning)
A technique where a previously trained AI model is used as the starting point for developing a new model for a different but typically related task – allowing knowledge gained in previous training cycles to be reapplied.
Transfer learning can drive efficiency savings by shortcutting model development. It can also be useful when data for the task that the model is being developed for is somewhat limited. But it’s important to note that the approach has limitations. Models that rely on transfer learning to gain generalized capabilities will likely require training on additional data in order to perform well in their domain of focus
(See: [Fine tuning](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#fine-tuning))
## [Weights](https://techcrunch.com/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/#weights)
Weights are core to AI training, as they determine how much importance (or weight) is given to different features (or input variables) in the data used for training the system — thereby shaping the AI model’s output.
Put another way, weights are numerical parameters that define what’s most salient in a dataset for the given training task. They achieve their function by applying multiplication to inputs. Model training typically begins with weights that are randomly assigned, but as the process unfolds, the weights adjust as the model seeks to arrive at an output that more closely matches the target.
For example, an AI model for predicting housing prices that’s trained on historical real estate data for a target location could include weights for features such as the number of bedrooms and bathrooms, whether a property is detached or semi-detached, whether it has parking, a garage, and so on.
Ultimately, the weights the model attaches to each of these inputs reflect how much they influence the value of a property, based on the given dataset.
*This article is updated regularly with new information.* |
| Shard | 190 (laksa) |
| Root Hash | 11303327329804289390 |
| Unparsed URL | com,techcrunch!/2023/07/10/black-sesame-ipo-nvidia-rival-auto-chip/ s443 |