Happy Monday! I spent this week keeping up with the latest AI / tech releases and updates so you don’t have to. Here are the 4 things I found most interesting.
OpenAI paused RL training on its newest models for two weeks after early signs one of them could cross its critical cyber threshold.
In an August 18 post, OpenAI said preliminary evidence that an upcoming model, Astra, may meet its Critical cybersecurity capability threshold pushed it to slow down, along with a security incident involving Hugging Face. That meant a two-week pause in reinforcement learning training on its newest models, and its largest planned frontier RL run was still on hold as of the post. Monitoring itself isn't cheap either, OpenAI puts the overhead at roughly 20% of the inference compute it watches.
Nvidia's agent hit a perfect 100 on ARC-AGI-3's public set, and the model underneath was Claude Opus 5.
On August 21 Nvidia said its agent system AVO cleared all 183 levels of the ARC-AGI-3 public set, scoring 100.00 on RHAE, the benchmark's human-efficiency metric. It ran on Claude Opus 5, which ARC Prize separately reports at about 30% on ARC-AGI-3 in its own harness, though Nvidia warns the two setups differ too much to call that a clean delta. These are public-set numbers only, not the private competition sets.
Anthropic hired a founder of Google's custom chip program, weeks after admitting it wants its own silicon.
Bloomberg reported on August 21 that Amir Salek is joining Anthropic's compute team. Per Bloomberg he was a founder of Alphabet's custom chip program and ran Google's TPU business until 2022, delivering the first seven generations of those chips. Anthropic had already confirmed on August 5 that it's building an in-house silicon team, and it has an initial order of roughly $250 million of chips from UK startup Fractile.
Wispr Flow hit a $2 billion valuation, roughly triple where it was nine months ago.
On August 17 the voice-to-text app announced a $280 million Series B led by Menlo Ventures, up from a $700 million valuation in November. It also previewed Canto, its first proprietary speech model, claiming word error rates in the hardest conditions fall from over 30% to somewhere between 5 and 10%. That's Wispr's own unbenchmarked number against its current model, and Canto hasn't shipped.
Those were the biggest AI updates on my radar this week.
On My Mind
One of the best skills you can adopt in any field is detaching yourself from results. Work you do with an expectation attached almost always comes out worse than work you do because you actually wanted to do it.
I started social media with absolutely zero expectations. I didn’t even know that this was an actual business. I just made things I wanted to make.
Then it grew, it became a serious contender to an actual job, and my behavior started to change. A video or an ad would perform bad and I’d get sad and stressed. In streaks it was even worse. And the thing is, the stuff I made without thinking about the algorithm almost always performed better than the stuff I made for it.
The second you make something for the result, you start making decisions differently. You take the safer idea. You edit toward whatever worked last time instead of what you actually wanted to say.
I still struggle with this a ton. Right now this is what I do for a living out of college, and I feel the anxiety of the instability this brings, so “just ignore the numbers” is easier said than done. But if I were to successfully detach myself from the results I believe I’d be in a much better place.
And content is just my example. A project, prepping for interviews, studying for exams, anything you pursue with the goal of getting great returns will probably perform worse than if you’d just let go. This is an insanely hard thing to achieve. But if we can manage to detach the work we put in from the results we get, all that’s left for us to do is the work.
More from me!
My last post: Top CS courses you can take for free! Take a look
Tool of the week: The voice tool I use every day to prompt AI (it's free and works on mobile too). Just clicking the link supports me! :)
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.
Have a great week, do more!
Volkan