New Reasoning Strategies Sweep Through OpenAI, Other Developers

Before we get to today’s column, we’re just three weeks away from AI Agenda Live, our flagship conference on all things artificial intelligence (a.k.a. my personal Super Bowl).

We’ll be joined in San Francisco by top researchers, founders, operators and investors who will be answering all of our burning questions about what’s coming up in bleeding-edge AI research, the next breakthrough AI use cases, and how AI startups are thinking about competing with their biggest model suppliers.

Speakers include Google DeepMind SVP & Chief AI Architect Koray Kavukcuoglu, who will be doing one of his first public interviews following the recent Google leadership shakeup, and OpenAI’s head of core products Thibault Sottiaux, who will be chatting about the ChatGPT maker’s product roadmap. You can’t miss it!

Onto today’s column…

Big things are coming from OpenAI with its forthcoming Astra model. As Amir, Rocket and I reported yesterday, the unreleased AI has some interesting new research powering it: specifically, an innovative technique that pushes the limits of how much the model is able to “think” about hard problems.

The breakthrough, related to well-known concepts such as loop transformers and recurrent depth, is causing all kinds of excitement as well as some measure of concern about how the technology could cause trouble.

On the performance front, some researchers believe the launch of Astra, with its stronger coding and computer-use abilities, could be akin to the performance jump when OpenAI launched GPT-4 in 2023. In other words, a much more impressive launch than when OpenAI launched GPT-5 last summer!

This would be welcome news for the AI industry as well as the global economy, given that the data center buildout is predicated on better AI that produces productivity gains and thus more revenues.

But there is a cause for concern, related to a shift in how the model works. Today, models reason about harder problems by writing out their thinking processes, otherwise known as “chains of thought.” This is helpful because researchers can monitor these chains of thought to make sure the models aren’t getting into any trouble (like trying to hack into other businesses in the case of Hugging Face and OpenAI).

The newer techniques allow models to “think” about hard problems by repeatedly running a question through the “layers” of mathematical operations that make up a model before producing the next word in an answer.

The result is that a model can end up displaying the capabilities of a much larger model, which is usually better at various tasks.

However, the model, depending on how much it leans into these new techniques, may not necessarily write out as much of its thinking processes the way it would with chains of thought. That means people may not be able to vet its thinking as easily.

That is worrying some researchers. After all, such chain-of-thought monitoring was one of the main solutions OpenAI put forward as a way to prevent a repeat of the Hugging Face hack in July.

And yet this technique is likely to spread. We’re hearing that these techniques are already a topic of conversation at every major lab, so we wouldn’t be surprised to see similar approaches at Anthropic, Google and others soon too.

To be fair, many researchers believe that chain-of-thought monitoring won’t be the final solution for how to monitor AI models for bad behavior. Researchers will have to come up with other monitoring techniques anyway. After all, models writing out their thought processes in this way is mostly a byproduct of how they’re trained today, and as they improve, it’s likely that models will tend to do less of this, researchers say, especially as model developers explore new optimizations and architectures.

For now, OpenAI is actually taking steps to ensure Astra’s chain of thought remains visible, even with this new technique, we hear. That could mean researchers are instructing the model to loop the question through the layer fewer times, for example. The fewer times a model loops a question through a layer of its mathematical computations, the less time it has to “think silently,” without revealing what steps or reasoning it is taking, researchers tell us.

Time will tell if new pressures will prompt other AI developers to discard such safeguards.—Stephanie Palazzolo, Rocket Drew and Amir Efrati

Here’s what else is going on…

Court Watch

On Tuesday, 30 new complaints were filed by Edelson PC, the law firm working on behalf of victims and families connected to the Tumbler Ridge mass shooting, against OpenAI. The new filings, for the first time, accuse OpenAI of aiding and abetting the mass shooting, rather than just negligently failing to prevent it.

Overheard

Elon Musk lamented a global “crisis of power” on Tuesday as the Tesla and SpaceX CEO spoke virtually at a G20 innovation event hosted in North Carolina by the White House Office of Science and Technology Policy and the Commerce Department. Musk cited a likely 15 gigawatt shortage of power next year for AI chips, calling on other countries to build data centers that could service AI companies desperate for computing capacity.

OpenAI plans to limit access to the most advanced cybersecurity capabilities of its forthcoming Astra model, according to a Tuesday blog post. The company said that it believes Astra meets the Critical cybersecurity threshold under its Preparedness Framework, meaning that the model can find previously unknown security flaws and ways to exploit those flaws across many well-protected systems without a person guiding each step. It is the first model the company is designating at this level, OpenAI said.

Meta Platforms’ chief AI officer, Alexandr Wang, told employees Monday that the company will move its internal communications from Google Chat to Slack later this month, according to an internal memo reviewed by The Information. Wang said that when Meta first evaluated Google Chat and Slack, the platforms offered relatively similar features, leading leadership to choose Google Chat because of its integration with Google Workspace. But less than a year after that decision, Meta is reversing course, with Wang acknowledging that another switch will be “disruptive” and “painful.”

Earnings Report

Palo Alto Networks lifted revenue 34% to $3.4 billion in the three months ending in July. CEO Nikesh Arora attributed the growth, which accelerated from 31% the quarter prior, to growing demand for cybersecurity from companies seeking to defend against new AI-powered threats. Still, Palo Alto Networks reported a net loss of $282 million in the quarter, compared with $254 net income a year prior, stemming in part from costs associated with its $25 billion acquisition of CyberArk earlier this year.

Dell said Tuesday that revenue for its quarter ending in July rose 58% from the year-ago quarter to $47 billion, beating the high end of its own forecast. The maker of personal computers and servers for AI data centers increased its outlook for the year ending in January 2027 by $25 billion to $192 billion, which would represent a 70% increase from the year-ago.

Deals and Debuts

See The Information’s Generative AI Database for an exclusive list of private companies and their investors.

AfterQuery, a data labeling company, is in talks to raise new funding at a $3.2 billion valuation, Forbes reported.

Nscale, which builds and runs data centers that rent out AI computing power, closed about $3 billion in loan facilities to fund two American sites—up to $1.85 billion for a campus in Ward County, Texas, and up to $1.2 billion for one in Madison, North Carolina.

Félix, whose service lets people in the United States send money to family in Latin America by chatting with an AI assistant inside WhatsApp, raised $200 million in a Series C funding round led by Andreessen Horowitz and General Catalyst.

HiBob, whose software manages payroll, hiring and employee records for mid-sized companies, received an investment led by Salesforce, with participation from Farallon. CTech reported $166 million raised at a valuation of about $3.2 billion, up from $2.7 billion in 2023.

Tripo AI, which builds AI models that generate three-dimensional objects and scenes, raised about 3 billion yuan ($450 million) across Series B and Series B+ funding rounds led by MPCi.

AIR, whose software blocks malicious instructions before an agent acts on tools and data they reach, emerged from stealth with $50 million led by Sequoia Capital and Greenoaks.

Light, whose service lets any business sell electricity to its own customers under its own brand by handling the licensing, buying and billing behind the scenes, raised $46 million in a Series A funding round led by Matrix.

Empirik, whose software watches for changes being made to a company's computer systems and predicts which ones will cause an outage, spun out of Sequoia Capital with $21 million in seed funding from Sequoia, Canapi and Alumni Ventures.

Aslan, which develops AI agents that work undercover in criminal forums and encrypted chat groups on behalf of the FBI and other agencies, launched publicly with $20.8 million led by Khosla Ventures and XYZ Venture Capital, according to Axios.

Norbert Health, whose software turns off-the-shelf robots into nursing assistants that take vital signs without touching a patient and write the results into the medical record, raised $14 million in a Series A funding round from William A. Marino with Cardinal Group.

Aranya, whose software takes a rack of bare servers and turns it into a working, self-repairing AI computing cluster in under two days, raised $11 million across two rounds: a $9 million seed funding round led by First Round Capital and a $2 million pre-seed funding round led by Asylum Ventures.

DataAgent, whose software finds and repairs failures inside a company's own cloud systems, emerged from stealth with $10 million in a pre-seed funding round led by MizMaa Ventures and Alicorn Venture Partners.

Visko, which offers AI models that generate an interactive world you can move around in as it streams, rather than a fixed video clip, raised $10 million in a pre-seed funding round led by Llama Ventures.

Xorlab, whose software learns what normal email traffic looks like inside a company so it can catch AI-written phishing messages that carry no malware, raised €5 million in a Series A+ funding round led by Spicehaus Partners.

Guickly, whose software shows a company every AI tool its employees are using and what each one costs, raised $4.2 million in a seed funding round led by Engineering Capital.

Palo Alto Networks has acquired Console, whose software lets a company describe an operational goal in plain English and then carries out the work automatically across its systems. Terms were not disclosed.

Salute, which runs and maintains data centers for their owners, has agreed to acquire T5 Operations, the facilities-management arm of T5 Data Centers. Terms were not disclosed.

SoftBank’s affiliate company SB Energy filed to go public, laying out in detail its ambition to become a neocloud. So far, however, as the filing showed, SB Energy has “no data center capacity” currently operating.

Anthropic released its latest models, Fable 5.1 and Mythos 5.1, on Tuesday. Mythos 5.1 will only be available to registered Anthropic partners engaged in either cybersecurity or life sciences research.

Meta Platforms on Tuesday unveiled a new audio transcription model, Muse Voice Transcribe, that CEO Mark Zuckerberg said can transcribe speech to text and segment audio based on who is speaking.

OpenAI on Tuesday announced that it would integrate ChatGPT Health with Epic’s electronic health record system to allow clinicians to import patient data and use AI to ask questions.

World Labs, a startup led by Stanford University professor Fei-Fei Li, on Tuesday unveiled Atlas, a new world model that can generate and modify photorealistic, interactive virtual environments based on text, image and video prompts.

Cerebras Systems and Compute Nordic Finland announced a new AI data center in Mikkeli, Finland, scaling in phases to 165 megawatts of contracted capacity, with construction on the first 50 megawatts already under way.

Thank you for reading the AI Agenda Newsletter! I’d love your feedback, ideas and tips: [email protected].

If you think someone else might enjoy this newsletter, please pass it forward or they can sign up here.

Stephanie Palazzolo is a reporter at The Information covering artificial intelligence. She previously worked at Business Insider covering AI and at Morgan Stanley as an investment banker. Based in New York, she can be reached at [email protected] or on Twitter at @steph_palazzolo.

Rocket Drew is a reporter at The Information covering AI. He can be reached at [email protected], on Signal at (530) 400-4184, or on Twitter @rocketalignment.

Amir Efrati is executive editor at The Information, which he helped to launch in 2013. Previously he spent nine years as a reporter at the Wall Street Journal, reporting on white-collar crime and later about technology. He can be reached at [email protected] and is on X @amir