What are people using AI models for? Despite the rapid growth in popularity of large language models, we still know very little about what they are used for.
This isn’t just out of curiosity, or even sociological research. Understanding how people actually use language models is important for security reasons: providers put a lot of effort into pre-deployment testing and use trust and safety systems to prevent misuse. But the scale and diversity of language models makes it hard to understand what they are used for (not to mention any kind of comprehensive security monitoring).
There’s another key factor that prevents us from having a clear understanding of how AI models are used: privacy. At Anthropic, our Claude model is not trained on user conversations by default, and we take protecting our users’ data very seriously. So how do we study and observe how our systems are used while strictly protecting our users’ privacy?
Cl aude Insights and Observations (“Clio” for short) is our attempt to answer this question. Clio is an automated analytics tool that performs privacy-preserving analysis of real-world language model usage. It gives us insight into everyday usage at claude.ai in a similar way to tools like Google Trends. It’s already helping us improve our security measures. In this post (with the full research paper attached), we describe Clio and some of its initial results.
How Clio Works: Privacy-Preserving Analytics at Scale
Traditional top-down security approaches (such as assessments and red teams) rely on knowing what to look for in advance. Clio takes a different approach, enabling bottom-up pattern discovery by distilling conversations into abstract, understandable clusters of topics. It does this while protecting user privacy: data is automatically anonymized and aggregated, and only higher-level clusters are visible to human analysts.
All of our privacy protections are extensively tested, as described in our research paper.
How People Use Claude: Insights from ClioUsing Clio, we were able to gain insight into how people use claude.ai in practice. While public datasets such as WildChat and LMSYS-Chat-1M provide useful information about how people use language models, they only capture specific contexts and use cases. Clio gives us a glimpse into the full real-world usage of claude.ai (which may differ from other AI systems due to differences in user base and model type).
Summary of Clio’s analysis steps, illustrated using fictional examples of conversations.
Here’s a brief overview of Clio’s multi-stage process:
Extracting Aspects: For each conversation, Clio extracts multiple “aspects” — specific properties or metadata, such as the topic of the conversation, the number of back-and-forths in the conversation, or the language used.
Semantic Clustering: Similar conversations are automatically grouped based on topics or general themes.
Cluster Descriptions: Each cluster receives a descriptive title and summary that captures common themes from the raw data while excluding private information.
Building Hierarchies: Clusters are organized into multi-level hierarchies for easier exploration. They can then be presented in an interactive interface that a human factors analyst can use to explore patterns along different dimensions (topics, language, etc.).
Shop Laptop Battery the Online Prices in UK – Large Selection of Batteries and Adapters from batteryforpc.co.uk With Best Prices
Category: news
Gli esperti di intelligenza artificiale rivelano il vero motivo per cui DeepSeek è così popolare
DeepSeek shocked the tech world last month. There’s a reason for that, according to AI experts, who say we’re likely just seeing the beginning of the Chinese tech startup’s influence in the field.
In late January, DeepSeek made headlines with its R1 AI model, which the company said roughly matched the performance of Open AI’s O1 model but cost a fraction of the price. DeepSeek briefly replaced ChatGPT as the top app in Apple’s App Store, sending tech stocks tumbling.
The achievement prompted U.S. tech giants to question America’s place in the AI race with China, and the billions of dollars behind those efforts. While Vice President JD Vance didn’t mention DeepSeek or China by name during his speech at the AI Action Summit in Paris on Tuesday, he did emphasize the importance of America’s lead in the field.
“The United States is a leader in AI , and our government plans to keep it that way,” he said, but added that “the United States wants to work with other countries.”
But there’s more to DeepSeek’s efficiency and capabilities than that. Experts say DeepSeek R1’s ability to reason and “think” answers to deliver high-quality results, combined with the company’s decision to make key parts of its technology public, will drive growth in the field.
While AI has long been used in tech products, it has reached a tipping point in the past two years thanks to the rise of ChatGPT and other generative AI services that have reshaped how people work, communicate and find information. It’s made companies like chipmaker Nvidia Wall Street darlings and upended the trajectory of Silicon Valley giants. So any development that helps build more powerful and efficient models is sure to be closely watched.
“This is definitely not hype,” said Oren Etzioni, former CEO of the Allen Institute for Artificial Intelligence. “But it’s also a world that’s changing very quickly.”
AI’s TikTok Moment
Tech leaders were quick to react to DeepSeek’s rise. Demis Hassabis, CEO of Google DeepMind, called the hype around DeepSeek “overblown,” but he also said the model was “probably the best work I’ve seen in China,” according to CNBC.
Microsoft CEO Satya Nadella said on the company’s quarterly earnings call in January that DeepSeek had some “real innovation,” while Apple CEO Tim Cook said on the iPhone maker’s earnings call that “innovation that drives efficiency is a good thing.”
But the attention isn’t all positive. Semiconductor research firm SemiAnalysis cast doubt on DeepSeek’s claim that it cost just $5.6 million to train. OpenAI told the Financial Times it found evidence that DeepSeek used the U.S. company’s models to train its own competitors.
“We are aware of and are reviewing indications that DeepSeek may have improperly improved our models, and we will share that information once we learn more,” an OpenAI spokesperson told CNN in a statement. DeepSeek was not immediately available for comment.
Two U.S. lawmakers called for a ban on the app on government devices after security researchers highlighted its possible links to the Chinese government, according to the Associated Press and ABC. Similar concerns have been raised about the popular social media app TikTok, which must be sold to a U.S. owner or risk being banned in the U.S.
“DeepSeek is the TikTok of (large language models),” Etzioni said.
How DeepSeek impressed the tech worldTech giants are already thinking about how DeepSeek’s technology will impact their products and services.
“DeepSeek basically gave us a solution in the form of a technical paper, but they didn’t provide the additional missing pieces,” said Lewis Tunstall, a senior research scientist at Hugging Face, an AI platform that provides tools for developers.
Tunstall is leading Hugging Face’s efforts to fully open source DeepSeek’s R1 model; while DeepSeek provided the research paper and model parameters, it did not reveal the code or training data.
Nadella said on Microsoft’s earnings call that Windows Copilot+ PCs (i.e. PCs built to specific specifications to support AI models) will be able to run AI models extracted from DeepSeek R1 locally. Mobile chip maker Qualcomm said on Tuesday that models extracted from DeepSeek R1 were running on smartphones and PCs equipped with its chips within a week.
AI researchers, academics and developers are still exploring what DeepSeek means for AI progress.
DeepSeek’s model isn’t the only open source model, nor is it the first that can reason about an answer before responding; OpenAI’s o1 model, which launched last year, can do that, too.
What makes DeepSeek so important is its ability to reason and learn from other models, and theAI community can see what’s going on behind the scenes. Those who use the R1 model in the DeepSeek app can also see how it “thinks” as it answers questions.
“You can see the wheels turning inside the machine,” Durga Malladi, senior vice president and general manager of technology planning and edge solutions at Qualcomm, told CNN.
iOS 18.3 is out, but AI is still waiting
An important milestone in the deployment of Apple intelligence in the United States, this iOS update still does not introduce AI on the European version of the iPhone.
We know what to expect, but impatience is the norm. While the American version of the iPhone can already benefit from numerous features related to artificial intelligence, in France, Apple intelligence will not show the end of its code until April at the latest. This does not prevent iOS 18.3, which has been launched on compatible smartphones since January 27, from offering some useful features. What does iOS 18.3 bring?
iOS 18.3 is not as harmful as it seems. At least in the United States, it brings the much-anticipated visual intelligence, Apple’s answer to Google Lens, which allows you to get information about objects, monuments, animals, just by putting them in the camera frame. Yes, it’s Pokémon. Too bad for us, we will have to wait a few months to find out.
The latest version of iOS is also the one that implements the deployment of Apple intelligence. Until now, Apple’s AI has been optional and enabled at the discretion of the user, now it will be enabled by default. To do this, iPhone users will have to go to Settings > Apple Intelligence and Siri.
In addition, the Genmoji image generation AIhas also been improved, and the news notification digest has been disabled. This is perhaps for the best, given the recent hype about the feature from Apple. What about us?Yes, everything described above is indeed reserved for iPhones on the other side of the Atlantic. For us Europeans, iOS 18.3 only fixes a few bugs. The calculator app has also restored its old functionality (and allows you to repeat actions by pressing the equals sign multiple times), and the keyboard no longer disappears when invoking Siri.
So, in our case, we can clearly talk about small updates. Anyway, who said that iOS 18.3 is finally public, also said that the iOS 18.4 beta is coming. And this time, it should be the right choice for European smartphone AI! We will be detailing the Apple Intelligence features that you can try in advance in this version in the coming weeks.
In the meantime, why not go and see what the iPhone SE 4 will look like?
ChatGPT just got OpenAI’s most powerful upgrade yet — and it’s ready for “deep research”
Artificial intelligence has already begun to change the way we conduct research, and now OpenAI has just released the latest update to ChatGPT, which it says will open up “deep research” to almost everyone.
The new deep research tool is rolling out to ChatGPT Pro subscribers today, just days after the launch of the 03-mini reasoning model.
“Today, we’re launching Deep Research in ChatGPT, a new agent capability that can conduct multi-step research on complex tasks on the internet,” a blog post reads, promising that the model can “do in just tens of minutes what would take a human hours to do.OpenAI ChatGPT’s powerful “deep research” upgrade gets open source copies in just 24 hours
I just tested ChatGPT’s new o3-mini model, rating its problem-solving and reasoning abilities with 7 prompts — and the results blew my mind
This new agent “can work independently for you,” “finding, analyzing, and synthesizing hundreds of online sources to create research analyst-level synthesis reports” based on a series of prompts.
It’s powered by the upcoming ChatGPT o3 model and is aimed at finance, science, policy, and engineering research, but OpenAI also says it can also be used to make purchasing decisions and help with detailed product research (which could be another warning for Google).
Using the new tool is simple, too. Users simply select “Drill down” in the ChatGPT message editor and enter a query, with the option to attach additional context via files and spreadsheets.
Doing so will see the cited articles in the sidebar, where you can also track your progress. OpenAI says it can take anywhere from five minutes to half an hour to complete a deep dive task.
Shop Laptop Battery the Online Prices in UK – Large Selection of Batteries and Adapters from batteryforpc.co.uk With Best Price
OpenAI launches new features for WhatsApp users — here’s what’s new
OpenAI has expanded the functionality of ChatGPT in WhatsApp to include uploading pictures, sending voice messages, and linking existing ChatGPT accounts directly through the messaging platform.
These updates have been rolled out to all users globally, ensuring that individuals around the world can benefit from the enhanced functionality. Users can now upload pictures within WhatsApp conversations using ChatGPT, similar to the functionality available in the standalone ChatGPT app. Users can now use AI to analyze and respond to visual content. ChatGPT is now on WhatsApp — here’s how to message the AI chatbot for free
You can now call or text 1-800-ChatGPT for free — here’s everything you need to knowIn addition, the integration of voice messages allows users to send voice notes to ChatGPT, which theAI will process and respond to in text form. While not exactly similar to ChatGPT Advanced Voice, this feature does provide a more natural interaction with ChatGPT within the platform, catering to a wider range of user preferences.
Account linking for expanded utilityIn addition, users now also have the option to link their existing ChatGPT account (whether Free, Plus, or Pro) to WhatsApp. This integration is designed to provide a more seamless experience, allowing users to manage their interactions and usage more effectively. By linking their accounts, users can enjoy extended usage and personalization, thereby enhancing the overall usefulness of ChatGPT in WhatsApp.
How to Use the New Features To use these new features, users should ensure that their WhatsApp application is updated to the latest version. Once updated, interacting with ChatGPT is very simple. Start by saving the number 1-800-CHATGPT (1-800-242-8478) to your phone contacts, then open WhatsApp and chat with your saved contacts to start a conversation. In the chat, use the attachment icon to select an image and send it directly to ChatGPT. To voice chat with ChatGPT, press and hold the microphone icon to record and send a voice message. ChatGPT will process these voice notes and reply in text.
If you want to get the best response, follow the prompts in the chat to link your ChatGPT account. You need to enable the extended usage and personalization features.
Benefits of the IntegrationThere are several benefits to integrating ChatGPT into WhatsApp. Accessing the AI assistant directly within the platform is much more convenient as it eliminates the need to switch between apps.
What is DeepSeek? Big tech companies continue to build AI in big ways.
About two weeks ago, Wall Street panicked when Chinese startup DeepSeek released an AI system that was far more effective than its American rivals.Investors, who have poured trillions of dollars into tech stocks over the past few years, wondered whether the tens of billions of dollars tech companies have spent on new data centers suddenly looked overextended.
But the biggest tech companies made clear in recent earnings reports that they don’t think they’re overextended in building new data centers.
Amazon hinted Thursday that its capital spending, which includes data center construction and other projects like warehouses, could exceed $100 billion this year. Microsoft said its spending could exceed $80 billion. Alphabet said it would spend $75 billion, while Meta reiterated plans for $65 billion in capital expenditures.
Combined, they could spend about $100 billion more on these projects than they did last year.
Executives urged patience. The problem, they said, is that customer demand for artificial intelligence outstrips the company’s capabilities. The only way to meet demand is to develop as many products as quickly as possible.
“Every time I see someone else doing something better, I say, ‘Ah, we should do that,’ ” Meta CEO Mark Zuckerberg told employees at a companywide meeting last week, according to a recording obtained by The New York Times. Competition is good, but we need to make sure we can win.”
Here are some keys to understanding this consumer-driven moment in the tech industry:
Tech companies need more data centers than they have now.
Many companies say they are constrained by the availability of chips, land and electricity to build data centers, so they are racing to open more. Microsoft, Alphabet and Amazon have all said their cloud sales could be higher if there was enough capacity. Cloud services are the quintessential way to deliver artificial intelligence to customers.
Alphabet Chief Financial Officer Anat Ashkenazi told investors that Alphabet is seeing “demand that exceeds our available capacity.” “So we’re going to work on that and make sure we can provide more capacity.”
Microsoft has been saying it has been constrained for some time, and previously told investors that pressure would ease early this year. But last week, when the company reported its latest earnings, executives told investors it might take until the summer to have enough capacity to meet all the demand. Shares of the company fell about 5% in after-hours trading after the report was released.
Greater efficiency, they say, will expand the use of and demand for AI
While many people think of data centers as expensive, power-hungry places to develop advanced AI systems, they are also where AI is deployed. Those are two different steps: training the model that underpins ChatGPT, and asking ChatGPT for recipe suggestions.
The industry calls deploying AI “inference”; a growing number of tech companies say their businesses are booming in this area.
Microsoft CEO Satya Nadella told investors last week that as costs fall, “AI I will become more pervasive.”
Amazon CEO Andy Jassy told investors Thursday that while a world where every application includes AI may be hard to imagine, “it’s a world we’ve been thinking about.” At the heart of that vision, he said, is inference.
He believes that lowering the cost of inference will follow the pattern of previous technology trends: As systems become cheaper to deploy, customers will “get excited about other things they can build that they previously thought were too expensive, and they usually end up spending more money on.”
These companies say they have to think long-term.
Cloud providers are used to giving customers the illusion of endless supply, which means they have to juggle having enough data centers online to play the video you want or answer your chatbot query. But they also can’t build too much in advance, which would tie up billions of dollars that could be deployed elsewhere. Balancing the two — especially when securing data center land, chips and electricity can take years — is a big challenge for these companies.
Executives say they can adjust how their investments are used, between building and deploying AI models and between serving their own core businesses and customers. Nadella said Microsoft’s infrastructure is “very flexible.” Ashkenazy said Google is flexible, too. For example, it can “repurpose capacity” to serve Google Search instead of cloud customers.
Zuckerberg said Meta was working on DeepSeek and how it could improve efficiency, but that investing heavily in data centers would be a strategic advantage against smaller, more nimble rivals.
“We serve over a billion people — that’s a lot of people, so more and more machines are going to be used to run inference,” he told employees.
Whatever the reason, cutting into profits — even the huge profits of tech giants — is unlikely to excite investors.
US reportedly investigating whether DeepSeek used restricted AI chips
The U.S. Commerce Department is investigating whether DeepSeek, a Chinese company whoseAI models have stunned the tech world with their performance, used chips banned from export by the U.S., according to people familiar with the matter.
China’s DeepSeek last week launched a free assistant that it said uses less data and costs a fraction of the U.S. model. Within days, it became the most downloaded app on Apple’s App Store and raised concerns about the U.S. lead in AI, sparking a sell-off that wiped out about $1 trillion in market value from U.S. tech stocks.
Current restrictions on Nvidia’s AI processors are aimed at blocking its most advanced chips from entering China.
Organized smuggling of AI chips into China has been traced to countries including Malaysia, Singapore and the United Arab Emirates, the sources said.
The U.S. Commerce Department and DeepSeek did not immediately respond to requests for comment.
An Nvidia spokesman said many of the company’s customers have business entities in Singapore and use them to make products for shipment to the U.S. and the West.
Nvidia said: “We insist that our partners comply with all applicable laws, and we will take appropriate action if we receive any information to the contrary.”
DeepSeek said it used Nvidia’s H800 chip, which can be legally purchased in 2023. Reuters could not determine whether DeepSeek used other regulated chips that are not allowed to be shipped to China.
DeepSeek also apparently has Nvidia’s lower-performance H20 chips, which can still be legally shipped to China. The United States has considered controlling these products during the Biden administration, and new Trump officials are also discussing it.
Dario Amodei, CEO of artificial intelligence company Anthropic, said earlier this week that “a significant portion of DeepSeek’s AI chip inventory appears to be chips that are not banned (but should be banned), chips that have been shipped before they were banned, and some chips that appear to be very likely smuggled.”
The United States has implemented a series of restrictions banning the export of AI chips to China and plans to restrict exports to many other countries.
Trump announces investment of up to $500 billion in private sector AI infrastructure
President Trump announced on Tuesday that the private sector will invest billions of dollars to build artificial intelligence infrastructure in the United States.
OpenAI, SoftBank and Oracle are planning to form a joint venture called “Stargate,” Trump said at a White House briefing.
SoftBank CEO Masayoshi Son, OpenAI’s Sam Altman and Oracle’s Larry Ellison joined Trump for the announcement.PoliticsTrump Announces Up to $500 Billion Investment in Private Sector AI Infrastructure
President Trump announced on Tuesday that the private sector will invest billions of dollars to build artificial intelligence infrastructure in the United States.
OpenAI, SoftBank and Oracle are planning to form a joint venture called “Stargate,” Trump said at a White House briefing.
SoftBank CEO Masayoshi Son, OpenAI’s Sam Altman and Oracle’s Larry Ellison joined Trump for the announcement.
US Politics TrumpPresident Donald Trump, SoftBank CEO Masayoshi Son and Open AI CEO Sam Altman listen to Oracle Executive Chairman Larry Ellison at the White House on January 21, 2025.
“What we want to do is, we want to keep these energy sources in this country,” Trump said. “China is a competitor, other countries are competitors. We want to stay in this country, and we are providing these energy sources. I will help a lot with the emergency declaration because we have an emergency, and we have to build these things. So they have to produce a lot of electricity. We will make it easy for them to do this in their own factories if they want.”
Executives from both companies are expected to invest $500 billion in the Stargate project over the next four years. Details of the new collaboration have not yet been announced.
Ellison said at the briefing that 10 data centers for the project are already under construction in Texas, and more are planned. Sources previously told CBS News that Stargate will start with a data center project in Texas and eventually expand to other states. AIholds incredible promise for all of us, for every American,” Ellison told reporters.
Trump claimed the venture would “create over 100,000 American jobs almost immediately.”
Altman added: “I think this is going to be the most important project of our time.”
The three executives touted what they believe is AI’s ability to solve health care problems.
Shop Laptop Battery the Online Prices in UK – Large Selection of Batteries and Adapters from batteryforpc.co.uk With Best Prices.
Cisco Launches AI Defense to Help Enterprises Transform with AI
Built for the enterprise, so enterprises can confidently develop, deploy, and secure AI applications.News Highlights:
Cisco’s end-to-end solution protects the development and use of AI applications, so enterprises can move forward with their AI initiatives with confidence.AI Defense protects against the misuse of AI tools, data breaches, and increasingly sophisticated threats that existing security solutions can’t handle.This innovative solution leverages Cisco’s unmatched network visibility and control to stay ahead of evolving AI security and safety issues.SAN JOSE, Calif., January 15, 2025 — Cisco (NASDAQ: CSCO), a leader in security and networking, today announced Cisco AI Defense, a breakthrough solution that enables and protects AI transformation within the enterprise. As AI advances, new security issues and security threats emerge at an unprecedented rate, and existing security solutions can’t keep up. Cisco AI Defense is built for enterprises, helping them confidently develop, deploy, and secure AI applications.
“When embracing AI, business and technology leaders cannot sacrifice security for speed. In a competitive, fast-changing environment, speed makes the difference. Built into the fabric of the network, Cisco AI Defense combines unique capabilities to detect and defend against threats as AI applications are developed and accessed, without the need to make trade-offs,” said Jeetu Patel, executive vice president and chief product officer at Cisco.
The risk of AI going wrong is extremely high. According to Cisco’s 2024 AI Readiness Index, only 29% of respondents believe they are fully capable of detecting and preventing unauthorized AI tampering. Because AI applications are multi-model and multi-cloud, the security challenges are also new and complex. Vulnerabilities can occur at the model or application level, and responsibility falls on different owners, including developers, end users, and vendors. As enterprises move beyond public data and begin training models on proprietary data, the risks only increase.
To unlock AI innovation and adoption, enterprises need a universal security layer to protect every user and every application. AI Defense supports the AI transformation of enterprises by addressing two pressing risks:
Develop and deploy secure AI applications: As AI becomes ubiquitous, enterprises will use and develop hundreds or even thousands of AI applications. Developers need a set of AI security safeguards that apply to each application. AI Defense helps developers move fast and unlock greater value by protecting AI systems from attacks and securing model behavior across platforms. AI Defense capabilities include:
Discover AI: Security teams need to understand who is building applications and what training sources they use. AI Defense detects shadow AI applications and sanctioned AIapplications in public and private clouds.
Model validation: Model tuning can lead to harmful and unexpected results. Automated testing checks AI models for hundreds of potential security issues. This AI-driven algorithmic red team identifies potential vulnerabilities and recommends guardrails in AI Defense for security teams to use.
Runtime security: Continuous validation provides ongoing protection against potential security threats such as tip injection, denial of service, and sensitive data leakage.Securing access to AI applications: As end users adopt AI applications such as summarization tools to increase productivity, security teams need to prevent data breaches and poisoning of proprietary data. AI Defense provides security teams with the following capabilities:
Visibility: Provides a comprehensive view of shadow and approved AI applications used by employees.
Access control: Enforces policies that limit employee access to unapproved AI tools.
Data and threat protection: Continuously protect against threats and loss of confidential data while ensuring compliance.
Unlike security guardrails built into individual AI models, Cisco provides consistent controls for a multi-model world. AI Defense is self-optimizing, leveraging Cisco’s proprietary machine learning models to detect evolving AI security issues based on threat intelligence data from Cisco Talos. Splunk customers using AI Defense will receive enriched alerts and more context from across the ecosystem. AI Defense seamlessly integrates with existing data flows, provides unparalleled visibility and control, and is built into Cisco’s unified AI-driven cross-domain security platform, Security Cloud. It leverages Cisco’s extensive network of enforcement points to enforce AIsecurity at the network level in a way that only Cisco can provide. Accuracy and trustworthiness are critical to protecting enterprise AI applications, and Cisco has been actively involved in setting industry standards for AI security, including those from MITRE, OWASP, and NIST.
Top 10 AI Predictions for 2025: AI Agents Will Go Mainstream
As 2024 draws to a close, venture capitalist Rob Toews from Radical Ventures shares his 10 predictions for AI in 2025:
01. Meta will start charging for Llama models
Meta is the world benchmark for open AI. In a compelling case study in corporate strategy, Meta has chosen to make its state-of-the-art Llama model available for free, while competitors like OpenAI and Google have closed sourced their cutting-edge models and charged for use.
So the news that Meta will start charging companies to use Llama next year will come as a surprise to many.
To be clear: we are not predicting that Meta will completely close source Llama, nor are we predicting that anyone using the Llama model will have to pay for it.
Rather, we are predicting that Meta will make the terms of Llama’s open source license more stringent so that companies above a certain size that use Llama in a commercial context will need to start paying to use the model.
Technically, Meta already does this today to a limited extent. The company doesn’t allow the largest companies — cloud supercomputers and other companies with more than 700 million monthly active users — to use its Llama models for free.
As early as 2023, Meta CEO Mark Zuckerberg said: “If you’re a company like Microsoft or Amazon or Google, and you’re basically reselling Llama, then we should get some revenue from it. I don’t think there will be a lot of revenue in the short term, but in the long term, hopefully there will be some revenue.”
Next year, Meta will significantly expand the range of companies that must pay to use Llama to include more large and medium-sized companies. Keeping up with the cutting edge of large language models (LLMs) is very expensive. Meta needs to invest billions of dollars each year to keep Llama consistent or close to consistent with the latest cutting-edge models from companies like OpenAI, Anthropic, and others.
Meta is one of the largest and best-funded companies in the world. But it’s also a public company and ultimately accountable to shareholders.
As the cost of making cutting-edge models continues to soar, it becomes increasingly untenable for Meta to invest so much money to train the next generation of Llama models without revenue expectations.
Over the next year, Llama models will continue to be available for free to enthusiasts, academics, individual developers, and startups. But 2025 will be the year Meta starts to seriously make Llama profitable.
02. Questions about “scaling laws”
In recent weeks, the most discussed topic in the field of artificial intelligence has been scaling laws, and the question of whether they are about to end.
Scaling laws were first proposed in an OpenAI paper in 2020. Its basic concept is simple and straightforward: when training an artificial intelligence model, as the number of model parameters, the amount of training data, and the amount of computation increase, the model’s performance will improve in a reliable and predictable way (technically, its test loss will decrease).
From GPT-2 to GPT-3 to GPT-4, the amazing performance gains are attributed to scaling laws.
Like Moore’s Law, scaling laws are not actually real laws, but just empirical observations.
In the past month, a series of reports have shown that major artificial intelligence labs are experiencing diminishing returns as large language models continue to scale up. This helps explain why OpenAI’s GPT-5 release has been repeatedly delayed.
The most common rebuttal to the stagnation of scaling laws is that the advent of test-time computation has opened up a whole new dimension in the pursuit of scaling.
That is, new inference models like OpenAI’s o3 can massively scale computation during inference, unlocking new AI capabilities by letting models “think longer”, rather than massively scaling computation during training.
This is an important point. Test-time computation does represent an exciting new avenue to achieve scaling and AI performance gains.
But another point about scaling laws is even more important, and one that is severely underappreciated in today’s discussion. Almost all discussions of scaling laws, starting with the original 2020 paper and continuing through today’s focus on test-time computation, have focused on language. But language is not the only data modality that matters.
Think about robotics, biology, world models, or networked agents. For these data modalities, scaling laws have not yet saturated; rather, they are just beginning.
In fact, rigorous proofs of scaling laws for these fields have not even been published to date.
Startups building foundational models for these new data patterns (e.g., evolutionary scaling in biology, physical intelligence in robotics, world labs in world models) are trying to identify and exploit scaling laws in these fields, just as OpenAI successfully exploited scaling laws for large language models (LLMs) in the first half of the 2020s.
Shop Laptop Battery the Online Prices in UK – Large Selection of Batteries and Adapters from batteryforpc.co.uk With Best Prices