The TechCrunch AI glossary | TechCrunch


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.


An AI agent refers to a tool that makes use of 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 different people can mean different things when they refer to an AI agent. Infrastructure is also still being built out to deliver on envisaged capabilities. But the basic concept implies an autonomous system that may draw on multiple AI systems to carry out multi-step tasks.

Given a simple question, a human brain can answer without even thinking too much about it — things like “which animal is taller between a giraffe and 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 right, especially in a logic or coding context. So-called reasoning models are developed from traditional large language models and optimized for chain-of-thought thinking thanks to reinforcement learning.

(See: Large language model)

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 AIs 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). It also typically takes longer to train deep learning vs. simpler machine learning algorithms — so development costs tend to be higher.

(See: Neural network)

This means further training of an AI model that’s intended 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 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))

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.

Those 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)

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 to take 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 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, whether for voice recognition, autonomous navigation, or drug discovery.

(See: Large language model (LLM))

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 data set 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 house 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, semi-detached, if it has or doesn’t have parking, a garage, and so on. 

Ultimately, the weights the model attaches to each of these inputs is a reflection of how much they influence the value of a property, based on the given data set.

DeepSeek: Everything you need to know about the AI chatbot app


DeepSeek has gone viral.

Chinese AI lab DeepSeek broke into the mainstream consciousness this week after its chatbot app rose to the top of the Apple App Store charts. DeepSeek’s AI models, which were trained using compute-efficient techniques, have led Wall Street analysts — and technologists — to question whether the U.S. can maintain its lead in the AI race and whether the demand for AI chips will sustain.

But where did DeepSeek come from, and how did it rise to international fame so quickly?

DeepSeek’s trader origins

DeepSeek is backed by High-Flyer Capital Management, a Chinese quantitative hedge fund that uses AI to inform its trading decisions.

AI enthusiast Liang Wenfeng co-founded High-Flyer in 2015. Wenfeng, who reportedly began dabbling in trading while a student at Zhejiang University, launched High-Flyer Capital Management as a hedge fund in 2019 focused on developing and deploying AI algorithms.

In 2023, High-Flyer started DeepSeek as a lab dedicated to researching AI tools separate from its financial business. With High-Flyer as one of its investors, the lab spun off into its own company, also called DeepSeek.

From day one, DeepSeek built its own datacenter clusters for model training. But like other AI companies in China, DeepSeek has been affected by U.S. export bans on hardware. To train one of its more recent models, the company was forced to use Nvidia H800 chips, a less-powerful version of a chip, the H100, available to U.S. companies.

DeepSeek’s technical team is said to skew young. The company reportedly aggressively recruits doctorate AI researchers from top Chinese universities. DeepSeek also hires people without any computer science background to help its tech better understand a wide range of subjects, per The New York Times.

DeepSeek’s strong models

DeepSeek unveiled its first set of models — DeepSeek Coder, DeepSeek LLM, and DeepSeek Chat — in November 2023. But it wasn’t until last spring, when the startup released its next-gen DeepSeek-V2 family of models, that the AI industry started to take notice.

DeepSeek-V2, a general-purpose text- and image-analyzing system, performed well in various AI benchmarks — and was far cheaper to run than comparable models at the time. It forced DeepSeek’s domestic competition, including ByteDance and Alibaba, to cut the usage prices for some of their models, and make others completely free.

DeepSeek-V3, launched in December 2024, only added to DeepSeek’s notoriety.

According to DeepSeek’s internal benchmark testing, DeepSeek V3 outperforms both downloadable, openly available models like Meta’s Llama and “closed” models that can only be accessed through an API, like OpenAI’s GPT-4o.

Equally impressive is DeepSeek’s R1 “reasoning” model. Released in January, DeepSeek claims R1 performs as well as OpenAI’s o1 model on key benchmarks.

Being a reasoning model, R1 effectively fact-checks itself, which helps it to avoid some of the pitfalls that normally trip up models. Reasoning models take a little longer — usually seconds to minutes longer — to arrive at solutions compared to a typical non-reasoning model. The upside is that they tend to be more reliable in domains such as physics, science, and math.

There is a downside to R1, DeepSeek V3, and DeepSeek’s other models, however. Being Chinese-developed AI, they’re subject to benchmarking by China’s internet regulator to ensure that its responses “embody core socialist values.” In DeepSeek’s chatbot app, for example, R1 won’t answer questions about Tiananmen Square or Taiwan’s autonomy.

A disruptive approach

If DeepSeek has a business model, it’s not clear what that model is, exactly. The company prices its products and services well below market value — and gives others away for free.

The way DeepSeek tells it, efficiency breakthroughs have enabled it to maintain extreme cost competitiveness. Some experts dispute the figures the company has supplied, however.

Whatever the case may be, developers have taken to DeepSeek’s models, which aren’t open source as the phrase is commonly understood but are available under permissive licenses that allow for commercial use. According to Clem Delangue, the CEO of Hugging Face, one of the platforms hosting DeepSeek’s models, developers on Hugging Face have created over 500 “derivative” models of R1 that have racked up 2.5 million downloads combined.

DeepSeek’s success against larger and more established rivals has been described as “upending AI” and ushering in “a new era of AI brinkmanship.” The company’s success was at least in part responsible for causing Nvidia’s stock price to drop by 18% on Monday, and for eliciting a public response from OpenAI CEO Sam Altman.

As for what DeepSeek’s future might hold, it’s not clear. Improved models are a given. But the U.S. government appears to be growing wary of what it perceives as harmful foreign influence.

TechCrunch has an AI-focused newsletter! Sign up here to get it in your inbox every Wednesday.

Open source companies that go proprietary: A timeline


Open source might be the building blocks of the modern software stack, but companies building businesses off the back of open source software face a perennial struggle between keeping their community happy and ensuring that third parties don’t abuse the permissions afforded by the license.

Many companies have launched with lofty open source ambitions, only to duck for cover once the realities of the commercial world hit home. It’s all about protecting their bottom line, especially with investors (public or private) to appease.

But it can be difficult keeping tabs on all these changes, while also distinguishing those that have abandoned open source altogether and those that have sought sanctuary behind a less permissive (but still open source) license (as the likes of Element and Grafana have done in the past few years).

As such, TechCrunch has compiled a timeline of open source companies that have changed course over the past decade.

Movable Type (2013)

Movable Type created an open source version (called MTOS) of its web publishing software in 2007 under a “copyleft” GPL open source license, a move that positioned it more closely to WordPress. Such licenses afford certain freedoms, but stipulate that all derivative work be released under a similar license. At any rate, this move lasted until 2013, at which point Movable Type’s then owners ditched the open source product, opining that it “hurt the adoption” of the commercial versions.

“The community has not grown because of MTOS, nor have we seen download numbers that are any greater than our paid versions of Movable Type, so at this point it does not make any economic sense to continue to maintain and distribute something that is getting very little use,” the company wrote at the time.

SugarCRM (2014)

Founded initially in 2004, customer relationship management (CRM) software maker SugarCRM announced in 2014 that it would no longer provide an open source “community edition,” noting that its two core markets — developers and first-time CRM users seeking a cheap solution — were not effectively being served by the product.

The company did continue to support the last version (v6.5) of the open source incarnation for four more years, before pulling the plug in 2018.

Redis (2018)

Redis, creators of the popular in-memory database store, has been transitioning away from its open source roots since 2018, when it moved its “Redis Modules” (e.g. RediSearch) from an open source AGPL license to Apache 2.0 with a “Commons Clause” addendum (i.e. commercial restrictions). The following year, Redis replaced the Commons Clause with its own Redis Source Available License (RSAL) that promised to maintain some freedoms, but with notable restrictions related to competing database services — such as those provided by companies such as AWS.

In many ways, this was a bellwether of what was to come, as other companies would later cite the “Amazon problem” as their reason for switching their license up. Earlier this year, Redis’ transition to the world of proprietary was complete, when it announced that its core software would be shifting from a BSD 3-Clause license to a dual-license setup — RSAL or server side public license (SSPL).

MongoDB (2018)

In 2018, database company MongoDB moved away from an open source AGPL license to SSPL. The reason? Yup: to prevent cloud hyperscalers such as AWS from selling their own version of the service without contributing back.

Confluent (2018)

The “year that was” for open source license switching concluded with Confluent, a company that sells enterprise-grade tools and services around Apache Kafka, switching some of the components of its core platform from Apache 2.0 to a proprietary Confluent Community License.

This license stipulates a notable exclusion, one that forbids any competing service from offering Confluent’s wares “as-a-service.”

Cockroach Labs (2019)

Cockroach Labs, creator of the eponymous distributed SQL database known as CockroachDB, has continued to shake up its licensing ethos.

In 2019, the company’s founders announced that they were moving CockroachDB from the permissive Apache 2.0 license to the Business Source License (BUSL). Again, cloud hyperscalers such as AWS were the driving force behind the change.

“We’re witnessing the rise of highly integrated providers take advantage of their unique position to offer ‘as-a-service’ versions of OSS [open source software] products, and offer a superior user experience as a consequence of their integrations,” the founders wrote at the time.

Back in August, Cockroach Labs announced yet another change: It would consolidate its self-hosted product under a single enterprise license, as a way to encourage larger businesses to pay for the features they really need. 

Sentry (2019)

Sentry, the $3 billion company behind the app performance monitoring platform of the same name, was once available under a permissive BSD 3-Clause open source license. But in 2019, the company moved to BUSL, with co-founder and CTO David Cramer saying this was to counter “funded businesses plagiarizing or copying our work to directly compete with Sentry.”

Last year, Sentry launched its very own Functional Source License (FSL), which is similar to BUSL but a little simpler. And as of this year, Sentry is putting its weight behind a new licensing paradigm dubbed “fair source,” which, as TechCrunch reported at the time, is “designed to bridge the open and proprietary worlds, replete with new definition, terminology, and governance model.”

Elastic (2021)

It was several years in the making, but Elastic — creator of enterprise search engine Elasticsearch and the Kibana visualization dashboard — went proprietary in 2021. It was a familiar story, one that can be traced back to 2015 when AWS launched its own managed Elasticsearch service.

However, Elastic stands somewhat alone as one of the only companies to move away from open source, and then move back. Back in August, Elastic announced it would be adopting an AGPL license — different to the Apache 2.0 license it used prior to 2021, but open source nonetheless.

HashiCorp (2023)

HashiCorp also abandoned the open source ship last year, announcing that it was switching its popular “infrastructure as code” tool Terraform from a copyleft open source license to BUSL.

The familiar reason was to prevent certain vendors from monetizing Terraform without contributing anything back to the project.

An open source fork called OpenTofu was launched earlier this year by third parties, and as a notable aside, IBM snapped up HashiCorp for $6.4 billion.

Snowplow (2024)

Snowplow, a VC-backed platform that helps companies collect behavioral data for AI applications, this year switched from an open source Apache 2.0 license to the Snowplow Limited Use License Agreement.

The reason, the company said, was that it needs to fund its “exciting technology roadmap,” and thus everyone running its software in production should “pay for the value they receive in return.” The new license also explicitly prevents users from creating a competitive product built on top of Snowplow.

The biggest data breaches in 2024: 1 billion stolen records and rising


We’re almost at the end of 2024, a year that will go down as having seen some of the biggest, most damaging data breaches in recent history. And just when you think that some of these hacks can’t get any worse, they do.

From huge stores of customers’ personal information getting scraped, stolen and posted online, to reams of medical data covering most people in the United States getting stolen, the worst data breaches of 2024 have surpassed the 1 billion stolen records and rising. These breaches not only affect the individuals whose data was irretrievably exposed, but also embolden the criminals who profit from their malicious cyberattacks.

Travel with us to the not-so-distant past to look at how some of the biggest security incidents of 2024 went down, their impact and, in some cases, how they could have been stopped. 

AT&T’s data breaches affect “nearly all” of its customers, and many more non-customers

For AT&T, 2024 has been a very bad year for data security. The telecoms giant confirmed not one, but two separate data breaches just months apart.

In July, AT&T said cybercriminals had stolen a cache of data that contained phone numbers and call records of “nearly all” of its customers, or around 110 million people, over a six-month period in 2022 and in some cases longer. The data wasn’t stolen directly from AT&T’s systems, but from an account it had with data giant Snowflake (more on that later).

Although the stolen AT&T data isn’t public (and one report suggests AT&T paid a ransom for the hackers to delete the stolen data) and the data itself does not contain the contents of calls or text messages, the “metadata” still reveals who called who and when, and in some cases the data can be used to infer approximate locations. Worse, the data includes phone numbers of non-customers who were called by AT&T customers during that time. That data becoming public could be dangerous for higher-risk individuals, such as domestic abuse survivors.

That was AT&T’s second data breach this year. Earlier in March, a data breach broker dumped online a full cache of 73 million customer records to a known cybercrime forum for anyone to see, some three years after a much smaller sample was teased online.

The published data included customers’ personal information, including names, phone numbers and postal addresses, with some customers confirming their data was accurate. 

But it wasn’t until a security researcher discovered that the exposed data contained encrypted passcodes used for accessing a customer’s AT&T account that the telecoms giant took action. The security researcher told TechCrunch at the time that the encrypted passcodes could be easily unscrambled, putting some 7.6 million existing AT&T customer accounts at risk of hijacks. AT&T force-reset its customers’ account passcodes after TechCrunch alerted the company to the researcher’s findings. 

One big mystery remains: AT&T still doesn’t know how the data leaked or where it came from. 

Change Healthcare hackers stole medical data on “substantial proportion” of people in America

In 2022, the U.S. Justice Department sued health insurance giant UnitedHealth Group to block its attempted acquisition of health tech giant Change Healthcare, fearing that the deal would give the healthcare conglomerate broad access to about “half of all Americans’ health insurance claims” each year. The bid to block the deal ultimately failed. Then, two years later, something far worse happened: Change Healthcare was hacked by a prolific ransomware gang; its almighty banks of sensitive health data were stolen because one of the company’s critical systems was not protected with multi-factor authentication.

The lengthy downtime caused by the cyberattack dragged on for weeks, causing widespread outages at hospitals, pharmacies and healthcare practices across the United States. But the aftermath of the data breach has yet to be fully realized, though the consequences for those affected are likely to be irreversible. UnitedHealth says the stolen data — which it paid the hackers to obtain a copy — includes the personal, medical and billing information on a “substantial proportion” of people in the United States. 

UnitedHealth has yet to attach a number to how many individuals were affected by the breach. The health giant’s chief executive, Andrew Witty, told lawmakers that the breach may affect around one-third of Americans, and potentially more. For now, it’s a question of just how many hundreds of millions of people in the U.S. are affected. 

Synnovis ransomware attack sparked widespread outages at hospitals across London 

A June cyberattack on U.K. pathology lab Synnovis — a blood and tissue testing lab for hospitals and health services across the U.K. capital — caused ongoing widespread disruption to patient services for weeks. The local National Health Service trusts that rely on the lab postponed thousands of operations and procedures following the hack, prompting the declaration of a critical incident across the U.K. health sector.

A Russia-based ransomware gang was blamed for the cyberattack, which saw the theft of data related to some 300 million patient interactions dating back a “significant number” of years. Much like the data breach at Change Healthcare, the ramifications for those affected are likely to be significant and life-lasting. 

Some of the data was already published online in an effort to extort the lab into paying a ransom. Synnovis reportedly refused to pay the hackers’ $50 million ransom, preventing the gang from profiting from the hack but leaving the U.K. government scrambling for a plan in case the hackers posted millions of health records online. 

One of the NHS trusts that runs five hospitals across London affected by the outages reportedly failed to meet the data security standards as required by the U.K. health service in the years that ran up to the June cyberattack on Synnovis.

Ticketmaster had an alleged 560 million records stolen in the Snowflake hack

A series of data thefts from cloud data giant Snowflake quickly snowballed into one of the biggest breaches of the year, thanks to the vast amounts of data stolen from its corporate customers. 

Cybercriminals swiped hundreds of millions of customer data from some of the world’s biggest companies — including an alleged 560 million records from Ticketmaster, 79 million records from Advance Auto Parts and some 30 million records from TEG — by using stolen credentials of data engineers with access to their employer’s Snowflake environments. For its part, Snowflake does not require (or enforce) its customers to use the security feature, which protects against intrusions that rely on stolen or reused passwords. 

Incident response firm Mandiant said around 165 Snowflake customers had data stolen from their accounts, in some cases a “significant volume of customer data.” Only a handful of the 165 companies have so far confirmed their environments were compromised, which also includes tens of thousands of employee records from Neiman Marcus and Santander Bank, and millions of records of students at Los Angeles Unified School District. Expect many Snowflake customers to come forward. 

(Dis)honorable mentions

Cencora notifies over a million and counting that it lost their data:

U.S. pharma giant Cencora disclosed a February data breach involving the compromise of patients’ health data, information that Cencora obtained through its partnerships with drug makers. Cencora has steadfastly refused to say how many people are affected, but a count by TechCrunch shows well over a million people have been notified so far. Cencora says it’s served more than 18 million patients to date. 

MediSecure data breach affects half of Australia:

Close to 13 million people in Australia — roughly half of the country’s population — had personal and health data stolen in a ransomware attack on prescriptions provider MediSecure in April. MediSecure, which distributed prescriptions for most Australians until late 2023, declared insolvency soon after the mass theft of customer data.

Kaiser shared health data on millions of patients with advertisers:

U.S. health insurance giant Kaiser disclosed a data breach in April after inadvertently sharing the private health information of 13.4 million patients, specifically website search terms about diagnoses and medications, with tech companies and advertisers. Kaiser said it used their tracking code for website analytics. The health insurance provider disclosed the incident in the wake of several  other telehealth startups, like Cerebral, Monument and Tempest, admitting they too shared data with advertisers.

USPS shared postal address with tech giants, too:

And then it was the turn of the U.S. Postal Service caught sharing postal addresses of logged-in users with advertisers like Meta, LinkedIn and Snap, using a similar tracking code provided by the companies. USPS removed the tracking code from its website after TechCrunch notified the postal service in July of the improper data sharing, but the agency wouldn’t say how many individuals had data collected. USPS has over 62 million Informed Delivery users as of March 2024.

Evolve Bank data breach affected fintech and startup customers:

A ransomware attack targeting Evolve Bank saw the personal information of more than 7.6 million people stolen by cybercriminals in July. Evolve is a banking-as-a-service giant serving mostly fintech companies and startups, like Affirm and Mercury. As a result, many of the individuals notified of the data breach had never heard of Evolve Bank, let alone have a relationship with the firm, prior to its cyberattack.

National Public Data goes broke after millions of SSNs stolen

The company behind the data broker National Public Data filed for Chapter 11 bankruptcy protection in October, months after a massive data breach exposed some three billion records affecting around 270 million individuals, according to various analyses by security researchers. The data broker allowed its paying customers access to its vast databases of names, dates of birth, email and postal addresses, phone numbers, and Social Security numbers (even if not all of the data was accurate). The company said it had to file for bankruptcy as it can no longer generate the revenue to address the deluge of class-action lawsuits and mounting liability from state and federal regulators.

First published on June 28 and updated on October 14.