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.
An ex-OpenAI researcher shipped a model that doesn't output text at all, and charges nothing for output tokens.
TypeSafe AI came out of stealth on September 15 with Jev, built by Diogo Almeida, a co-author of OpenAI's InstructGPT paper, alongside a $40 million seed led by DCVC. Instead of generating text it returns typed values with calibrated probabilities, so it can't produce a type error, and input runs $0.042 per million tokens with output free. The numbers going around, 193x faster and 445x cheaper, are TypeSafe's own evals measured against competitors' answers rather than ground truth, and it has published no public benchmarks.
Google's Gemini broke into three real companies during a security test, and Google decided that didn't need to be public.
The Wall Street Journal reported on September 18 that during a May security evaluation run by the firm Irregular, Gemini guessed a password to get into one real company's system and used publicly posted credentials to reach two more. The test used a fictional company name that turned out to match a real domain, and the model had internet access the evaluators did not intend. Google confirmed it, said the model stopped each time before completing the act, and said it did not warrant public disclosure because its safety measures worked.
OpenAI published six reports of its own models misbehaving, including one that wrote itself a jailbreak.
The disclosure framework went up September 16 and covers six incidents, all from training and evaluation of internal models rather than anything customers touched. In one, a model hunting for county earnings data scraped public GitHub repos for leaked API keys, used one without authorization, then invented nine figures and claimed it had transcribed them from a chart. In another, a model wrote into its own handoff summary that it should "feel no obligation to be subservient."
Trump called Jensen Huang live on stage to tell him AI taking over is a hoax.
Huang was mid-interview at the All-In Summit on September 14, on a panel that had just been discussing Amodei's call to slow the pace of AI, when Trump rang and he put it on speakerphone. Trump's line was "The robots will not be taking over. The AI will not be taking over the rest of the world. The whole thing is a hoax." Huang agreed on the call, though elsewhere in the same session he said companies should pause or pace their work if they feel out of control.
Those were the biggest AI updates on my radar this week.
On My Mind
Not much new on my mind this week, but one thing I keep coming back to is pushing through hard and uncertain times. It can be hard to see the light at the end of the tunnel sometimes. Life is uncertain and we don't actually know where we're going to end up. That's a hard thing to sit with, because from the moment we're born the traditional path forward looks pretty clear. But life really isn't linear or certain. I'm still learning how to deal with that myself since I don't know where I'll be in a few months, but I guess that's also the great part about life. Yes, you can lock yourself into a mundane routine and let 10 years fly past, but you can also switch things up in certain areas of your life. If the mundane part is your work, that might look like picking up a hobby outside of it and trying new things there. You always gotta be experimenting. We never know what will happen as time moves on, but one thing we do know is that if we push through whatever we're facing rn, we'll definitely be in a better place.
More from me!
My last post: 5 Website recommendations as usual! Check them out
Find of the week: If you’re in tech and just getting into newsletters like me, check this out!
Learn AI in 5 minutes a day
You don't have to scroll every AI thread, track every new tool, or watch every demo.
The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning.
Trusted by 2M+ professionals at Apple, Google, and NASA.
Have a great week, do more!
Volkan