Sam Altman and Dario Amodei Warn the AI Race Is Getting Riskier

Morning, {{first name | folks}}! A lot happened in AI over the weekend. The big labs are talking safety, xAI pushed back Grok 4.7, OpenAI is rebuilding infrastructure for a billion users, DeepSeek is coming after inference costs, and researchers are finding new uses for AI in healthcare.

Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have both raised concerns about how quickly AI capabilities are advancing. Amodei warned that the next 6–12 months could bring much more capable AI agents and said the industry may need to slow down if safety systems can't keep pace. Altman similarly said OpenAI would pause development if it believed its models could no longer be built safely, adding that even a 10% chance of a catastrophic outcome would be unacceptable.

Altman's interview also touched on OpenAI's next phase as a company. He said the company is now targeting an IPO in 2027, while acknowledging that going public could create pressure to move faster at exactly the time when safety may require more caution. That tension is becoming harder to ignore as the leading labs push toward more autonomous systems.

xAI has delayed Grok 4.7 after testing exposed a problem with how the model handles difficult tasks. Elon Musk said it was sometimes giving up on problems it could solve and wasn't doing enough work to check its own answers. The release, which was expected around September 11, is now being pushed back while xAI works on the issue. Musk said the model needs to be more persistent and rigorous before it's ready to ship.

For a model being positioned as xAI's next big jump, that's notable. The company isn't just chasing benchmark scores here; it's trying to make the model spend more effort when a problem calls for it. That kind of behavior could matter a lot more once models are doing longer coding and agent tasks on their own.

OpenAI rebuilt its Habitat storage system as ChatGPT usage grew to more than 1 billion weekly users. It now handles over 70 million requests per second across nearly 40 regions and stores more than 500 petabytes of data. The company also rewrote most of Habitat from Python to Rust, with two engineers using Codex and GPT-5.5 to help with the rewrite.

At that scale, the problem isn't just handling more traffic. OpenAI found that it also had to control how teams use the system, putting limits on queries that could create problems for everyone else. It's a good reminder that when infrastructure gets this big, the rules around using it can matter just as much as the infrastructure itself.

DeepSeek released V4.1-Flash, an open-weight model aimed at coding and agent workloads with a 1-million-token context window and support for text and image input. The model uses a mixture-of-experts setup with 552 billion total parameters but activates only a fraction of them for each request, helping keep inference costs down. DeepSeek has also made it available under the MIT license.

The bigger move is on pricing. DeepSeek says V4.1-Flash can handle work that previously went to V4 Pro, and from September 14, Pro requests are being routed to Flash and charged at the lower Flash rate. For teams running large AI workloads, cheaper inference can matter just as much as another jump in benchmark scores.

Researchers are testing AI models that analyze how people speak to spot patterns linked to schizophrenia. In one study, a model looking at 88 speech features reached 86.2% accuracy, while another that looked at the meaning and flow of conversations reached 87%, compared with 68% for clinical raters.

The reason researchers are interested is timing. It can take around 18 months for someone to receive a diagnosis after symptoms first appear, so picking up subtle changes earlier could make a difference. The models still need to be tested on much larger and more diverse groups before they can be used in clinical settings, but speech is turning out to be a surprisingly useful source of information.

Other AI Signals: