There’s a lot going on in the world of Mac apps for popular AI services. In the past week, Anthropic has released a desktop app for its popular Claude chatbot, and Perplexity launched a native app for its AI-driven search service.
On top of that, OpenAI updated its ChatGPT Mac app with support for its flashy advanced voice feature.
Like the ChatGPT app that debuted several weeks ago, the Perplexity app adds a keyboard shortcut that allows you to enter a query from anywhere on your desktop. You can use the app to ask follow-up questions and carry on a conversation about what it finds.
It’s free to download and use, but Perplexity offers subscriptions for major users.
Perplexity’s search emphasis meant it wasn’t previously a direct competitor to OpenAI’s ChatGPT, but OpenAI recently launched SearchGPT, a search-focused variant of its popular product. SearchGPT is not yet supported in the desktop app, though.
Anthropic’s Claude, on the other hand, is a more direct competitor to ChatGPT. It works similarly to ChatGPT but has different strengths, particularly in software development. The Claude app is free to download, but it’s in beta, and like Perplexity and OpenAI, Anthropic charges for more advanced users.
When ChatGPT launched its Mac app, it didn’t release a Windows app right away, saying that it was focused on where its users were at the time. A Windows app recently arrived, and Anthropic took a different approach, simultaneously introducing Windows and Mac apps.
Previously, all these tools offered mobile apps and web apps, but not necessarily native desktop apps.
The large language model-based coding assistant GitHub Copilot will switch from using exclusively OpenAI’s GPT models to a multi-model approach over the coming weeks, GitHub CEO Thomas Dohmke announced in a post on GitHub’s blog.
First, Anthropic’s Claude 3.5 Sonnet will roll out to Copilot Chat’s web and VS Code interfaces over the next few weeks. Google’s Gemini 1.5 Pro will come a bit later.
Additionally, GitHub will soon add support for a wider range of OpenAI models, including GPT o1-preview and o1-mini, which are intended to be stronger at advanced reasoning than GPT-4, which Copilot has used until now. Developers will be able to switch between the models (even mid-conversation) to tailor the model to fit their needs—and organizations will be able to choose which models will be usable by team members.
The new approach makes sense for users, as certain models are better at certain languages or types of tasks.
“There is no one model to rule every scenario,” wrote Dohmke. “It is clear the next phase of AI code generation will not only be defined by multi-model functionality, but by multi-model choice.”
It starts with the web-based and VS Code Copilot Chat interfaces, but it won’t stop there. “From Copilot Workspace to multi-file editing to code review, security autofix, and the CLI, we will bring multi-model choice across many of GitHub Copilot’s surface areas and functions soon,” Dohmke wrote.
There are a handful of additional changes coming to GitHub Copilot, too, including extensions, the ability to manipulate multiple files at once from a chat with VS Code, and a preview of Xcode support.
GitHub Spark promises natural language app development
In addition to the Copilot changes, GitHub announced Spark, a natural language tool for developing apps. Non-coders will be able to use a series of natural language prompts to create simple apps, while coders will be able to tweak more precisely as they go. In either use case, you’ll be able to take a conversational approach, requesting changes and iterating as you go, and comparing different iterations.
Competing companies like OpenAI have been working on equivalent tools but have not made them publicly available yet. It’s something of an arms race, as these tools are projected to generate a lot of revenue in a few years if they progress as expected.
There’s a belief that these tools could eventually automate many menial tasks in office jobs. It could also be a useful tool for developers in that it could “automate repetitive tasks” and streamline laborious QA and optimization work.
That has long been part of Anthropic’s message to investors: Its AI tools could handle large portions of some office jobs more efficiently and affordably than humans can. The public testing of the Computer Use feature is a step toward achieving that goal.
We’re, of course, familiar with the ongoing argument about these types of tools between the “it’s just a tool that will make people’s jobs easier” and the “it will put people out of work across industries like a wrecking ball”—both of these things could happen to some degree. It’s just a question of what the ratio will be—and that may vary by situation or industry.
There are numerous valid concerns about the widespread deployment of this technology, though. To its credit, Anthropic has tried to anticipate some of these by putting safeguards in from the get-go. The company gave some examples in its blog post:
Our teams have developed classifiers and other methods to flag and mitigate these kinds of abuses. Given the upcoming US elections, we’re on high alert for attempted misuses that could be perceived as undermining public trust in electoral processes. While computer use is not sufficiently advanced or capable of operating at a scale that would present heightened risks relative to existing capabilities, we’ve put in place measures to monitor when Claude is asked to engage in election-related activity, as well as systems for nudging Claude away from activities like generating and posting content on social media, registering web domains, or interacting with government websites.
These safeguards may not be perfect, as there may be creative ways to circumvent them or other unintended consequences or misuses yet to be discovered.
Right now, Anthropic is putting Computer Use out there for testing to see what problems arise and to work with developers to improve its capabilities and find positive uses.
Coding shortcuts in canvas include reviewing code, adding logs for debugging, inserting comments, fixing bugs, and porting code to different programming languages. For example, if your code is JavaScript, with a few clicks it can become PHP, TypeScript, Python, C++, or Java. As with GPT-4o by itself, you’ll probably still have to check it for mistakes.
A screenshot of coding using ChatGPT with Canvas captured on October 4, 2024.
Credit: Benj Edwards
A screenshot of coding using ChatGPT with Canvas captured on October 4, 2024. Credit: Benj Edwards
Also, users can highlight specific sections to direct ChatGPT’s focus, and the AI model can provide inline feedback and suggestions while considering the entire project, much like a copy editor or code reviewer. And the interface makes it easy to restore previous versions of a working document using a back button in the Canvas interface.
A new AI model
OpenAI says its research team developed new core behaviors for GPT-4o to support Canvas, including triggering the canvas for appropriate tasks, generating certain content types, making targeted edits, rewriting documents, and providing inline critique.
An image of OpenAI’s Canvas in action.
An image of OpenAI’s Canvas in action. Credit: OpenAI
One key challenge in development, according to OpenAI, was defining when to trigger a canvas. In an example on the Canvas blog post, the team says it taught the model to open a canvas for prompts like “Write a blog post about the history of coffee beans” while avoiding triggering Canvas for general Q&A tasks like “Help me cook a new recipe for dinner.”
Another challenge involved tuning the model’s editing behavior once canvas was triggered, specifically deciding between targeted edits and full rewrites. The team trained the model to perform targeted edits when users specifically select text through the interface, otherwise favoring rewrites.
The company noted that canvas represents the first major update to ChatGPT’s visual interface since its launch two years ago. While canvas is still in early beta, OpenAI plans to improve its capabilities based on user feedback over time.
OpenAI and Anthropic have each signed unprecedented deals granting the US government early access to conduct safety testing on the companies’ flashiest new AI models before they’re released to the public.
According to a press release from the National Institute of Standards and Technology (NIST), the deal creates a “formal collaboration on AI safety research, testing, and evaluation with both Anthropic and OpenAI” and the US Artificial Intelligence Safety Institute.
Through the deal, the US AI Safety Institute will “receive access to major new models from each company prior to and following their public release.” This will ensure that public safety won’t depend exclusively on how the companies “evaluate capabilities and safety risks, as well as methods to mitigate those risks,” NIST said, but also on collaborative research with the US government.
The US AI Safety Institute will also be collaborating with the UK AI Safety Institute when examining models to flag potential safety risks. Both groups will provide feedback to OpenAI and Anthropic “on potential safety improvements to their models.”
NIST said that the agreements also build on voluntary AI safety commitments that AI companies made to the Biden administration to evaluate models to detect risks.
Elizabeth Kelly, director of the US AI Safety Institute, called the agreements “an important milestone” to “help responsibly steward the future of AI.”
Anthropic co-founder: AI safety “crucial” to innovation
Among the most controversial aspects of the bill is a requirement that AI companies build in a “kill switch” to stop models from introducing “novel threats to public safety and security,” especially if the model is acting “with limited human oversight, intervention, or supervision.”
Critics say the bill overlooks existing safety risks from AI—like deepfakes and election misinformation—to prioritize prevention of doomsday scenarios and could stifle AI innovation while providing little security today. They’ve urged California’s governor, Gavin Newsom, to veto the bill if it arrives at his desk, but it’s still unclear if Newsom intends to sign.
Anthropic was one of the AI companies that cautiously supported California’s controversial AI bill, Reuters reported, claiming that the potential benefits of the regulations likely outweigh the costs after a late round of amendments.
The company’s CEO, Dario Amodei, told Newsom why Anthropic supports the bill now in a letter last week, Reuters reported. He wrote that although Anthropic isn’t certain about aspects of the bill that “seem concerning or ambiguous,” Anthropic’s “initial concerns about the bill potentially hindering innovation due to the rapidly evolving nature of the field have been greatly reduced” by recent changes to the bill.
In a letter to the bill’s co-sponsor, California Senator Scott Wiener, OpenAI’s chief strategy officer, Jason Kwon, suggested that “the federal government should lead in regulating frontier AI models to account for implications to national security and competitiveness.”
The ChatGPT maker striking a deal with the US AI Safety Institute seems in line with that thinking. As Kwon told Reuters, “We believe the institute has a critical role to play in defining US leadership in responsibly developing artificial intelligence and hope that our work together offers a framework that the rest of the world can build on.”
While some critics worry California’s AI safety bill will hamper innovation, Anthropic’s co-founder, Jack Clark, told Reuters today that “safe, trustworthy AI is crucial for the technology’s positive impact.” He confirmed that Anthropic’s “collaboration with the US AI Safety Institute” will leverage the government’s “wide expertise to rigorously test” Anthropic’s models “before widespread deployment.”
In NIST’s press release, Kelly agreed that “safety is essential to fueling breakthrough technological innovation.”
By directly collaborating with OpenAI and Anthropic, the US AI Safety Institute also plans to conduct its own research to help “advance the science of AI safety,” Kelly said.
The United Kingdom’s Competition and Markets Authority (CMA) has officially launched a probe into Amazon’s $4 billion partnership with the AI firm Anthropic, as it continues to monitor how the largest tech companies might seize control of AI to further entrench their dominant market positions.
Through the partnership, “Amazon will become Anthropic’s primary cloud provider for certain workloads, including agreements for purchasing computing capacity and non-exclusive commitments to make Anthropic models available on Amazon Bedrock,” the CMA said.
Amazon and Anthropic deny there’s anything wrong with the deal. But because the CMA has seen “some” foundational model (FM) developers “form partnerships with major cloud providers” to “secure access to compute” needed to develop models, the CMA is worried that “incumbent firms” like Amazon “could use control over access to compute to shape FM-related markets in their own interests.”
Due to this potential risk, the CMA said it is “considering” whether Amazon’s partnership with Anthropic “has resulted in the creation of a relevant merger situation under the merger provisions of the Enterprise Act 2002 and, if so, whether the creation of that situation has resulted, or may be expected to result, in a substantial lessening of competition within any market or markets” in the UK.
It’s not clear yet if Amazon’s partnership with Anthropic is problematic, but the CMA confirmed that after a comment period last April, it now has “sufficient information” to kick off this first phase of its merger investigation.
By October 4, this first phase will conclude, after which the CMA may find that the partnership does not qualify as a merger situation, the UK regulator said. Or it may determine that it is a merger situation “but does not raise competition concerns,” clearing Amazon to proceed with the deal.
However, if a merger situation exists, and “it may result in a substantial lessening of competition” in a UK market, the CMA may refer the investigation to the next phase, allowing a panel of independent experts to dig deeper to illuminate potential risks and concerns. If Amazon wants to avoid that deeper probe potentially ordering steep fines, the tech giant would then have the option to offer fixes to “resolve the CMA’s concerns,” the CMA said.
An Amazon spokesperson told Reuters that its “collaboration with Anthropic does not raise any competition concerns or meet the CMA’s own threshold for review.”
“Amazon holds no board seat nor decision-making power at Anthropic, and Anthropic is free to work with any other provider (and indeed has multiple partners),” Amazon’s spokesperson said, defending the deal.
Anthropic’s spokesperson agreed that nothing was amiss, telling Reuters that “our strategic partnerships and investor relationships do not diminish our corporate governance independence or our freedom to partner with others. We intend to cooperate with the CMA and provide them with a comprehensive understanding of Amazon’s investment and our commercial collaboration.”
In the AI world, there’s a buzz in the air about a new AI language model released Tuesday by Meta: Llama 3.1 405B. The reason? It’s potentially the first time anyone can download a GPT-4-class large language model (LLM) for free and run it on their own hardware. You’ll still need some beefy hardware: Meta says it can run on a “single server node,” which isn’t desktop PC-grade equipment. But it’s a provocative shot across the bow of “closed” AI model vendors such as OpenAI and Anthropic.
“Llama 3.1 405B is the first openly available model that rivals the top AI models when it comes to state-of-the-art capabilities in general knowledge, steerability, math, tool use, and multilingual translation,” says Meta. Company CEO Mark Zuckerberg calls 405B “the first frontier-level open source AI model.”
In the AI industry, “frontier model” is a term for an AI system designed to push the boundaries of current capabilities. In this case, Meta is positioning 405B among the likes of the industry’s top AI models, such as OpenAI’s GPT-4o, Claude’s 3.5 Sonnet, and Google Gemini 1.5 Pro.
A chart published by Meta suggests that 405B gets very close to matching the performance of GPT-4 Turbo, GPT-4o, and Claude 3.5 Sonnet in benchmarks like MMLU (undergraduate level knowledge), GSM8K (grade school math), and HumanEval (coding).
But as we’ve noted many times since March, these benchmarks aren’t necessarily scientifically sound or translate to the subjective experience of interacting with AI language models. In fact, this traditional slate of AI benchmarks is so generally useless to laypeople that even Meta’s PR department now just posts a few images of charts and doesn’t even try to explain them in any detail.
Enlarge/ A Meta-provided chart that shows Llama 3.1 405B benchmark results versus other major AI models.
We’ve instead found that measuring the subjective experience of using a conversational AI model (through what might be called “vibemarking”) on A/B leaderboards like Chatbot Arena is a better way to judge new LLMs. In the absence of Chatbot Arena data, Meta has provided the results of its own human evaluations of 405B’s outputs that seem to show Meta’s new model holding its own against GPT-4 Turbo and Claude 3.5 Sonnet.
Enlarge/ A Meta-provided chart that shows how humans rated Llama 3.1 405B’s outputs compared to GPT-4 Turbo, GPT-4o, and Claude 3.5 Sonnet in its own studies.
Whatever the benchmarks, early word on the street (after the model leaked on 4chan yesterday) seems to match the claim that 405B is roughly equivalent to GPT-4. It took a lot of expensive computer training time to get there—and money, of which the social media giant has plenty to burn. Meta trained the 405B model on over 15 trillion tokens of training data scraped from the web (then parsed, filtered, and annotated by Llama 2), using more than 16,000 H100 GPUs.
So what’s with the 405B name? In this case, “405B” means 405 billion parameters, and parameters are numerical values that store trained information in a neural network. More parameters translate to a larger neural network powering the AI model, which generally (but not always) means more capability, such as better ability to make contextual connections between concepts. But larger-parameter models have a tradeoff in needing more computing power (AKA “compute”) to run.
We’ve been expecting the release of a 400 billion-plus parameter model of the Llama 3 family since Meta gave word that it was training one in April, and today’s announcement isn’t just about the biggest member of the Llama 3 family: There’s an entirely new iteration of improved Llama models with the designation “Llama 3.1.” That includes upgraded versions of its smaller 8B and 70B models, which now feature multilingual support and an extended context length of 128,000 tokens (the “context length” is roughly the working memory capacity of the model, and “tokens” are chunks of data used by LLMs to process information).
Meta says that 405B is useful for long-form text summarization, multilingual conversational agents, and coding assistants and for creating synthetic data used to train future AI language models. Notably, that last use-case—allowing developers to use outputs from Llama models to improve other AI models—is now officially supported by Meta’s Llama 3.1 license for the first time.
Abusing the term “open source”
Llama 3.1 405B is an open-weights model, which means anyone can download the trained neural network files and run them or fine-tune them. That directly challenges a business model where companies like OpenAI keep the weights to themselves and instead monetize the model through subscription wrappers like ChatGPT or charge for access by the token through an API.
Fighting the “closed” AI model is a big deal to Mark Zuckerberg, who simultaneously released a 2,300-word manifesto today on why the company believes in open releases of AI models, titled, “Open Source AI Is the Path Forward.” More on the terminology in a minute. But briefly, he writes about the need for customizable AI models that offer user control and encourage better data security, higher cost-efficiency, and better future-proofing, as opposed to vendor-locked solutions.
All that sounds reasonable, but undermining your competitors using a model subsidized by a social media war chest is also an efficient way to play spoiler in a market where you might not always win with the most cutting-edge tech. That benefits Meta, Zuckerberg says, because he doesn’t want to get locked into a system where companies like his have to pay a toll to access AI capabilities, drawing comparisons to “taxes” Apple levies on developers through its App Store.
Enlarge/ A screenshot of Mark Zuckerberg’s essay, “Open Source AI Is the Path Forward,” published on July 23, 2024.
So, about that “open source” term. As we first wrote in an update to our Llama 2 launch article a year ago, “open source” has a very particular meaning that has traditionally been defined by the Open Source Initiative. The AI industry has not yet settled on terminology for AI model releases that ship either code or weights with restrictions (such as Llama 3.1) or that ship without providing training data. We’ve been calling these releases “open weights” instead.
Unfortunately for terminology sticklers, Zuckerberg has now baked the erroneous “open source” label into the title of his potentially historic aforementioned essay on open AI releases, so fighting for the correct term in AI may be a losing battle. Still, his usage annoys people like independent AI researcher Simon Willison, who likes Zuckerberg’s essay otherwise.
“I see Zuck’s prominent misuse of ‘open source’ as a small-scale act of cultural vandalism,” Willison told Ars Technica. “Open source should have an agreed meaning. Abusing the term weakens that meaning which makes the term less generally useful, because if someone says ‘it’s open source,’ that no longer tells me anything useful. I have to then dig in and figure out what they’re actually talking about.”
The Llama 3.1 models are available for download through Meta’s own website and on Hugging Face. They both require providing contact information and agreeing to a license and an acceptable use policy, which means that Meta can technically legally pull the rug out from under your use of Llama 3.1 or its outputs at any time.
On Thursday, Anthropic announced Claude 3.5 Sonnet, its latest AI language model and the first in a new series of “3.5” models that build upon Claude 3, launched in March. Claude 3.5 can compose text, analyze data, and write code. It features a 200,000 token context window and is available now on the Claude website and through an API. Anthropic also introduced Artifacts, a new feature in the Claude interface that shows related work documents in a dedicated window.
So far, people outside of Anthropic seem impressed. “This model is really, really good,” wrote independent AI researcher Simon Willison on X. “I think this is the new best overall model (and both faster and half the price of Opus, similar to the GPT-4 Turbo to GPT-4o jump).”
As we’ve written before, benchmarks for large language models (LLMs) are troublesome because they can be cherry-picked and often do not capture the feel and nuance of using a machine to generate outputs on almost any conceivable topic. But according to Anthropic, Claude 3.5 Sonnet matches or outperforms competitor models like GPT-4o and Gemini 1.5 Pro on certain benchmarks like MMLU (undergraduate level knowledge), GSM8K (grade school math), and HumanEval (coding).
Enlarge/ Claude 3.5 Sonnet benchmarks provided by Anthropic.
If all that makes your eyes glaze over, that’s OK; it’s meaningful to researchers but mostly marketing to everyone else. A more useful performance metric comes from what we might call “vibemarks” (coined here first!) which are subjective, non-rigorous aggregate feelings measured by competitive usage on sites like LMSYS’s Chatbot Arena. The Claude 3.5 Sonnet model is currently under evaluation there, and it’s too soon to say how well it will fare.
Claude 3.5 Sonnet also outperforms Anthropic’s previous-best model (Claude 3 Opus) on benchmarks measuring “reasoning,” math skills, general knowledge, and coding abilities. For example, the model demonstrated strong performance in an internal coding evaluation, solving 64 percent of problems compared to 38 percent for Claude 3 Opus.
Claude 3.5 Sonnet is also a multimodal AI model that accepts visual input in the form of images, and the new model is reportedly excellent at a battery of visual comprehension tests.
Enlarge/ Claude 3.5 Sonnet benchmarks provided by Anthropic.
Roughly speaking, the visual benchmarks mean that 3.5 Sonnet is better at pulling information from images than previous models. For example, you can show it a picture of a rabbit wearing a football helmet, and the model knows it’s a rabbit wearing a football helmet and can talk about it. That’s fun for tech demos, but the tech is still not accurate enough for applications of the tech where reliability is mission critical.
On Thursday, DuckDuckGo unveiled a new “AI Chat” service that allows users to converse with four mid-range large language models (LLMs) from OpenAI, Anthropic, Meta, and Mistral in an interface similar to ChatGPT while attempting to preserve privacy and anonymity. While the AI models involved can output inaccurate information readily, the site allows users to test different mid-range LLMs without having to install anything or sign up for an account.
DuckDuckGo’s AI Chat currently features access to OpenAI’s GPT-3.5 Turbo, Anthropic’s Claude 3 Haiku, and two open source models, Meta’s Llama 3 and Mistral’s Mixtral 8x7B. The service is currently free to use within daily limits. Users can access AI Chat through the DuckDuckGo search engine, direct links to the site, or by using “!ai” or “!chat” shortcuts in the search field. AI Chat can also be disabled in the site’s settings for users with accounts.
According to DuckDuckGo, chats on the service are anonymized, with metadata and IP address removed to prevent tracing back to individuals. The company states that chats are not used for AI model training, citing its privacy policy and terms of use.
“We have agreements in place with all model providers to ensure that any saved chats are completely deleted by the providers within 30 days,” says DuckDuckGo, “and that none of the chats made on our platform can be used to train or improve the models.”
Enlarge/ An example of DuckDuckGo AI Chat with GPT-3.5 answering a silly question in an inaccurate way.
Benj Edwards
However, the privacy experience is not bulletproof because, in the case of GPT-3.5 and Claude Haiku, DuckDuckGo is required to send a user’s inputs to remote servers for processing over the Internet. Given certain inputs (i.e., “Hey, GPT, my name is Bob, and I live on Main Street, and I just murdered Bill”), a user could still potentially be identified if such an extreme need arose.
While the service appears to work well for us, there’s a question about its utility. For example, while GPT-3.5 initially wowed people when it launched with ChatGPT in 2022, it also confabulated a lot—and it still does. GPT-4 was the first major LLM to get confabulations under control to a point where the bot became more reasonably useful for some tasks (though this itself is a controversial point), but that more capable model isn’t present in DuckDuckGo’s AI Chat. Also missing are similar GPT-4-level models like Claude Opus or Google’s Gemini Ultra, likely because they are far more expensive to run. DuckDuckGo says it may roll out paid plans in the future, and those may include higher daily usage limits or access to “more advanced models.”)
It’s true that the other three models generally (and subjectively) pass GPT-3.5 in capability for coding with lower hallucinations, but they can still make things up, too. With DuckDuckGo AI Chat as it stands, the company is left with a chatbot novelty with a decent interface and the promise that your conversations with it will remain private. But what use are fully private AI conversations if they are full of errors?
Enlarge/ Mixtral 8x7B on DuckDuckGo AI Chat when asked about the author. Everything in red boxes is sadly incorrect, but it provides an interesting fantasy scenario. It’s a good example of an LLM plausibly filling gaps between concepts that are underrepresented in its training data, called confabulation. For the record, Llama 3 gives a more accurate answer.
Benj Edwards
As DuckDuckGo itself states in its privacy policy, “By its very nature, AI Chat generates text with limited information. As such, Outputs that appear complete or accurate because of their detail or specificity may not be. For example, AI Chat cannot dynamically retrieve information and so Outputs may be outdated. You should not rely on any Output without verifying its contents using other sources, especially for professional advice (like medical, financial, or legal advice).”
So, have fun talking to bots, but tread carefully. They’ll easily “lie” to your face because they don’t understand what they are saying and are tuned to output statistically plausible information, not factual references.
With most computer programs—even complex ones—you can meticulously trace through the code and memory usage to figure out why that program generates any specific behavior or output. That’s generally not true in the field of generative AI, where the non-interpretable neural networks underlying these models make it hard for even experts to figure out precisely why they often confabulate information, for instance.
Now, new research from Anthropic offers a new window into what’s going on inside the Claude LLM’s “black box.” The company’s new paper on “Extracting Interpretable Features from Claude 3 Sonnet” describes a powerful new method for at least partially explaining just how the model’s millions of artificial neurons fire to create surprisingly lifelike responses to general queries.
Opening the hood
When analyzing an LLM, it’s trivial to see which specific artificial neurons are activated in response to any particular query. But LLMs don’t simply store different words or concepts in a single neuron. Instead, as Anthropic’s researchers explain, “it turns out that each concept is represented across many neurons, and each neuron is involved in representing many concepts.”
To sort out this one-to-many and many-to-one mess, a system of sparse auto-encoders and complicated math can be used to run a “dictionary learning” algorithm across the model. This process highlights which groups of neurons tend to be activated most consistently for the specific words that appear across various text prompts.
Enlarge/ The same internal LLM “feature” describes the Golden Gate Bridge in multiple languages and modes.
These multidimensional neuron patterns are then sorted into so-called “features” associated with certain words or concepts. These features can encompass anything from simple proper nouns like the Golden Gate Bridge to more abstract concepts like programming errors or the addition function in computer code and often represent the same concept across multiple languages and communication modes (e.g., text and images).
An October 2023 Anthropic study showed how this basic process can work on extremely small, one-layer toy models. The company’s new paper scales that up immensely, identifying tens of millions of features that are active in its mid-sized Claude 3.0 Sonnet model. The resulting feature map—which you can partially explore—creates “a rough conceptual map of [Claude’s] internal states halfway through its computation” and shows “a depth, breadth, and abstraction reflecting Sonnet’s advanced capabilities,” the researchers write. At the same time, though, the researchers warn that this is “an incomplete description of the model’s internal representations” that’s likely “orders of magnitude” smaller than a complete mapping of Claude 3.
Enlarge/ A simplified map shows some of the concepts that are “near” the “inner conflict” feature in Anthropic’s Claude model.
Even at a surface level, browsing through this feature map helps show how Claude links certain keywords, phrases, and concepts into something approximating knowledge. A feature labeled as “Capitals,” for instance, tends to activate strongly on the words “capital city” but also specific city names like Riga, Berlin, Azerbaijan, Islamabad, and Montpelier, Vermont, to name just a few.
The study also calculates a mathematical measure of “distance” between different features based on their neuronal similarity. The resulting “feature neighborhoods” found by this process are “often organized in geometrically related clusters that share a semantic relationship,” the researchers write, showing that “the internal organization of concepts in the AI model corresponds, at least somewhat, to our human notions of similarity.” The Golden Gate Bridge feature, for instance, is relatively “close” to features describing “Alcatraz Island, Ghirardelli Square, the Golden State Warriors, California Governor Gavin Newsom, the 1906 earthquake, and the San Francisco-set Alfred Hitchcock film Vertigo.”
Enlarge/ Some of the most important features involved in answering a query about the capital of Kobe Bryant’s team’s state.
Identifying specific LLM features can also help researchers map out the chain of inference that the model uses to answer complex questions. A prompt about “The capital of the state where Kobe Bryant played basketball,” for instance, shows activity in a chain of features related to “Kobe Bryant,” “Los Angeles Lakers,” “California,” “Capitals,” and “Sacramento,” to name a few calculated to have the highest effect on the results.
Enlarge/ The Claude AI iOS app running on an iPhone.
Anthropic
On Wednesday, Anthropic announced the launch of an iOS mobile app for its Claude 3 AI language models that are similar to OpenAI’s ChatGPT. It also introduced a new subscription tier designed for group collaboration. Before the app launch, Claude was only available through a website, an API, and other apps that integrated Claude through API.
Like the ChatGPT app, Claude’s new mobile app serves as a gateway to chatbot interactions, and it also allows uploading photos for analysis. While it’s only available on Apple devices for now, Anthropic says that an Android app is coming soon.
Anthropic rolled out the Claude 3 large language model (LLM) family in March, featuring three different model sizes: Claude Opus, Claude Sonnet, and Claude Haiku. Currently, the app utilizes Sonnet for regular users and Opus for Pro users.
While Anthropic has been a key player in the AI field for several years, it’s entering the mobile space after many of its competitors have already established footprints on mobile platforms. OpenAI released its ChatGPT app for iOS in May 2023, with an Android version arriving two months later. Microsoft released a Copilot iOS app in January. Google Gemini is available through the Google app on iPhone.
Enlarge/ Screenshots of the Claude AI iOS app running on an iPhone.
Anthropic
The app is freely available to all users of Claude, including those using the free version, subscribers paying $20 per month for Claude Pro, and members of the newly introduced Claude Team plan. Conversation history is saved and shared between the web app version of Claude and the mobile app version after logging in.
Speaking of that Team plan, it’s designed for groups of at least five and is priced at $30 per seat per month. It offers more chat queries (higher rate limits), access to all three Claude models, and a larger context window (200K tokens) for processing lengthy documents or maintaining detailed conversations. It also includes group admin tools and billing management, and users can easily switch between Pro and Team plans.
On Friday, the US Department of Homeland Security announced the formation of an Artificial Intelligence Safety and Security Board that consists of 22 members pulled from the tech industry, government, academia, and civil rights organizations. But given the nebulous nature of the term “AI,” which can apply to a broad spectrum of computer technology, it’s unclear if this group will even be able to agree on what exactly they are safeguarding us from.
President Biden directed DHS Secretary Alejandro Mayorkas to establish the board, which will meet for the first time in early May and subsequently on a quarterly basis.
The fundamental assumption posed by the board’s existence, and reflected in Biden’s AI executive order from October, is that AI is an inherently risky technology and that American citizens and businesses need to be protected from its misuse. Along those lines, the goal of the group is to help guard against foreign adversaries using AI to disrupt US infrastructure; develop recommendations to ensure the safe adoption of AI tech into transportation, energy, and Internet services; foster cross-sector collaboration between government and businesses; and create a forum where AI leaders to share information on AI security risks with the DHS.
It’s worth noting that the ill-defined nature of the term “Artificial Intelligence” does the new board no favors regarding scope and focus. AI can mean many different things: It can power a chatbot, fly an airplane, control the ghosts in Pac-Man, regulate the temperature of a nuclear reactor, or play a great game of chess. It can be all those things and more, and since many of those applications of AI work very differently, there’s no guarantee any two people on the board will be thinking about the same type of AI.
This confusion is reflected in the quotes provided by the DHS press release from new board members, some of whom are already talking about different types of AI. While OpenAI, Microsoft, and Anthropic are monetizing generative AI systems like ChatGPT based on large language models (LLMs), Ed Bastian, the CEO of Delta Air Lines, refers to entirely different classes of machine learning when he says, “By driving innovative tools like crew resourcing and turbulence prediction, AI is already making significant contributions to the reliability of our nation’s air travel system.”
So, defining the scope of what AI exactly means—and which applications of AI are new or dangerous—might be one of the key challenges for the new board.
A roundtable of Big Tech CEOs attracts criticism
For the inaugural meeting of the AI Safety and Security Board, the DHS selected a tech industry-heavy group, populated with CEOs of four major AI vendors (Sam Altman of OpenAI, Satya Nadella of Microsoft, Sundar Pichai of Alphabet, and Dario Amodei of Anthopic), CEO Jensen Huang of top AI chipmaker Nvidia, and representatives from other major tech companies like IBM, Adobe, Amazon, Cisco, and AMD. There are also reps from big aerospace and aviation: Northrop Grumman and Delta Air Lines.
Upon reading the announcement, some critics took issue with the board composition. On LinkedIn, founder of The Distributed AI Research Institute (DAIR) Timnit Gebru especially criticized OpenAI’s presence on the board and wrote, “I’ve now seen the full list and it is hilarious. Foxes guarding the hen house is an understatement.”
