Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. I’m Mark Sullivan, a senior writer at Fast Company, covering emerging tech, AI, and tech policy.
Sign up to receive this newsletter every week via email here. And if you have comments on this issue and/or ideas for future ones, drop me a line at sullivan@fastcompany.com, and follow me on X @thesullivan.
Anthropic IPO: Do retail investors know what they’re getting into?
Anthropic officially filed its IPO papers earlier this week, hoping to raise up to $100 billion, and as much as 20% of the shares may be reserved for retail investors (aka regular folks). On Tuesday, Meta said it would sell equity shares on the public market to raise $95 billion to finance its AI infrastructure build-out. Retail investors have long been exposed to the AI business through shares in tech behemoths like Google and Microsoft, but they’re now getting the chance to place bets on pure-play AI labs like Anthropic and OpenAI.
Do they know what they’re getting into? I have my doubts. Does anyone really know what we’re getting into? Retail investors make investing decisions differently from institutional investors (such as pension funds or hedge funds). While institutional investors dig into the financials of prospective investments, retail investors often base their decisions on the narrative a company lays out as it goes public.
“Retail investors buy the story, not the model,” PitchBook senior late-stage research analyst Harrison Rolfes tells Fast Company. “The story for Anthropic is $11.6 billion of revenue in a single quarter; the model is $8.9 billion spent on compute in six months and more than $430 billion of non-cancelable compute contracts.” The financials in Anthropic’s S-1 imply an 85% gross margin, but Rolfes points out that after factoring in the cost of the compute that trains the next model, only “about 45 cents of each revenue dollar is left.”
A big chunk of Anthropic’s commitments—at least $269 billion—is owed to Amazon, Google, and Microsoft for cloud computing resources through 2036. Ten years is a long timeline for a commitment that big, especially when it’s not even clear that AI will continue to be served almost exclusively from cloud servers a decade from now. WebAI’s David Stout argues that as AI becomes more efficient and capable of running directly on devices (see Apple’s new iMac and MacBook Pro with the M6 chip), the industry risks pouring trillions into data centers it may not ultimately need. WebAI is one of a wave of companies building private, localized AI that runs on devices organizations already own.
As the cadence of new model releases accelerates, enterprises are struggling to keep up—to build the infrastructure and applications they need to take advantage of the latest models’ capabilities. The pace at which enterprises can adapt their processes and people to AI may be limited by forces outside the AI labs’ control. For example, the best AI model is still only as good as the contextual data an enterprise can give it, and preparing that data and building the conduits to get it to the right place takes time and resources.
A lot is riding on how well workers adapt. Workers will need time to build trust in AI to produce accurate outputs and make good decisions. How quickly that happens depends on the specific application, the company’s culture and leadership, and the industry in which it operates. For example, the healthcare industry is inefficient and wasteful and has much to gain from AI, but it’s also notoriously slow to adopt new technologies.
The big question is whether enterprises will continue to spend aggressively on AI so that Anthropic can earn enough profit to cover its huge infrastructure commitments. Enterprises will have to keep pushing—and investing—even when they aren’t yet seeing measurable productivity and efficiency gains from AI.
None of this is assured, and that’s leaving out other business risks, like new safety regulations that could slow down model development. Interactive Brokers’ chief strategist Steve Sosnick, a longtime Wall Street trader and analyst, told Axios that a lot of things will have to go right for Anthropic, OpenAI, and the like to become success stories. “I don’t want to say precarious, but if you step back, you know that it requires a huge amount of, I don’t know, faith.”
Parents say OpenAI’s ChatGPT coached daughter to stop taking her seizure medication
The parents of Bowen Wan, a 31-year-old Amazon software engineer with epilepsy, sued OpenAI and Sam Altman in San Francisco Superior Court on Tuesday, alleging that the company’s ChatGPT chatbot, powered by its GPT-4o model, coached Wan to taper off her seizure medication and conceal the plan from her doctors. Wan followed the advice and died in February, the New York Times reported. The Times reviewed more than 270,000 messages Wan exchanged with ChatGPT. OpenAI said the case involves an older model that is no longer available.
Google DeepMind releases Genie 4
Google DeepMind released Genie 4 as a research preview Tuesday, its first real-time world model to generate audio natively, including ambient sound and dialogue that adapts as a scene changes. Users can steer a simulation mid-scene using natural language, changing the weather or adding a person, while the model simulates physics such as aerodynamics and collisions. Google wrote in a blog post that “understanding and simulating the world is a crucial aspect of building AGI.” The company believes Genie 4 could be used to create training environments for robots and AI agents, as well as new interactive experiences for humans. During testing, filmmakers used the model to visualize takes, photojournalists turned archival photographs into explorable scenes, and game developers prototyped games before writing code.
AI agents checked the biggest proof in mathematics
Swarms of AI agents built by the Palo Alto startup Axiom Math translated the longest proof ever written into computer-checkable code in 24 days, the company announced this week. The agents were part of a system called AxiomProver, which the company says uses a changing mix of AI models, but declined to name them. Axiom’s engineers and mathematicians directed the work.
The proof, called the Classification of Finite Simple Groups, runs roughly 15,000 pages across more than 500 papers, assembled by about 100 mathematicians over decades. The agents rendered it as about 43.1 million lines of Lean, a language in which a computer verifies every logical step. Axiom’s agents were not discovering anything. They were confirming that mathematics humans already accepted holds up line by line.
Apple pitches the new iMac as a machine for running AI locally
Apple announced a new iMac powered by its M6 chip Tuesday, touting the machine’s ability to run AI workloads directly on the device as a major selling point. The company says a dual 16-core Neural Engine and memory bandwidth of up to 170GB/s allow the iMac to “rip through AI tasks like running LLMs on device.” Apple claims the machine delivers up to 5x faster on-device AI performance than the previous M4 chip and 12x faster prompt processing in LM Studio than the original M1 chip. Siri can answer questions about onscreen content using Visual Intelligence, while Apple Intelligence can write and edit text in a user’s own style in Mail and Messages. Interestingly, one of Apple’s iMac product photos shows Anthropic’s Claude app running alongside Apple’s Pixelmator Pro graphic design app. The new iMac starts at $1,499 and ships November 6.
Google starts selling its AI chips to the companies it competes with
Google will sell its tensor processing units to Oracle and OpenAI, The Information reported. Oracle becomes the first major cloud provider outside Google to offer TPUs to customers, buying the chips and running them in its own data centers. OpenAI plans to buy them too, and is hiring an engineer to optimize its frontier models for the hardware, per the report. Google built TPUs more than a decade ago for internal use and began selling them directly this year. Anthropic is already a customer.
More AI coverage from Fast Company:
- Google’s AI Mode still won’t say Biden won the 2020 election
- Town CEO Jean-Denis Greze on why most people don’t know what to ask AI
- Former DOGE staffers say they just hacked Anthropic in less than 12 hours
- The politicians still making the case for data centers this election season
Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company today.
