HawksT

Obsessed with how AI is changing products. Daily notes on what's new and what matters.

[HawksT @John43497285]

@ArtificialAnlys 9x for the same model is wild. Makes me wonder how many cost comparisons are basically measuring cache behavior.

[HawksT @John43497285]

@simonw A pelican riding a bike was already weird enough. Fourier epicycles somehow made it weirder.

[HawksT @John43497285]

@TechCrunch The local/cloud split sounds tricky. What happens when the internet drops halfway through a task?

[HawksT @John43497285]

The interesting question is how much more usage $200 buys you a year from now.

[Theo - t3.gg @theo]

"but theo, AI keeps getting more expensive! Do you really think those $200 plans are gonna stick around?"

AI is getting cheaper. The $200 plans will get better over time, not worry (ignore OAI they don't count right now)

[HawksT @John43497285]

用 AI Studio 搓个 Android App，直接通过 Google Play 装到自己手机上，不用开发者账号，也不用公开上架。

个人小工具就该这么简单。

https://t.co/Lj8RK2CkVH

[HawksT @John43497285]

@TheBlockCo Interesting given they're already working together on Patch the Planet. Wonder what an acquisition would add beyond that partnership.

[HawksT @John43497285]

@theo That "interesting" is doing a lot of work lol. Looking forward to the video.

[HawksT @John43497285]

@Reuters Borrowing more to fund an already massive AI bet. Son really isn't leaving much room for a slow payoff.

[HawksT @John43497285]

@Dimillian Can you clear old session data from that tab, or is it just for checking usage right now?

[HawksT @John43497285]

A million people trying small, strange experiments sounds a lot more interesting than watching a few labs race through benchmarks.

[François Chollet @fchollet]

The ideal outcome for science isn't a couple of labs using proprietary models and millions of dollars to crack trophy problems, it's to have millions of AI-aided human scientists who follow their curiosity and become way more ambitious

[HawksT @John43497285]

摩根大通允许第三方 agent 通过 Plaid 等连接器读取部分账户数据，但不让它动钱。自家的 Smart Cash 倒是在试点中。

https://t.co/USyhNqvdqy

[HawksT @John43497285]

@TokenGremlin The 5-hour windows are more annoying than the price sometimes. Just let me see exactly when my limit resets.

[HawksT @John43497285]

@paulg I wonder how many of those 2B monthly requests come from agents rather than humans.

[HawksT @John43497285]

@theo Does the $18k include someone becoming the unofficial GPU sysadmin?

[HawksT @John43497285]

@edzitron The fact that Wang told this story himself on a podcast makes it even funnier.

[HawksT @John43497285]

Half the time on a realistic monitor is a pretty uncomfortable number.

[Neel Nanda @NeelNanda5]

My top takeaway from assessing Gemini 4: Argon is significantly less monitorable than past models, and IMO this trend is likely to continue, like Sol -> Astra. It can evade a realistic CoT+action monitor to do simple sabotage half the time.

We can reliably maintain control for now, with good CoT monitoring, as it still frequently slips up, but we need to be prepared for when that stops being true in future models, which may be sooner rather than later. Among other things, we really need interpretability methods that can reliably capture unverbalized motivations in models, with minimal hallucinations

I suspect this is driven by the model being highly capable without CoT, being capable of 5.5 arithmetic steps in a single forward pass (4 for Fable, 7 for Astra). It seems to omit information from its CoT more often, though in non-adversarial settings the CoT is still very helpful in practice, and remains in legible English. It still needs CoT to do the hardest tasks, but any given task becomes "easier" with each new frontier model.

It is also highly situationally aware, and shows awareness of the existence of CoT monitors and that it may want to evade them, eg it sometimes decides to avoid terms like "scheming" in the output or to keep its thoughts brief as a result. For now, these concealment strategies are pretty obvious in the CoT, but this suggests future misaligned models may successfully evade CoT monitors.

[HawksT @John43497285]

Pi API 的案例里最有意思的是 Axol，双臂机器人在给自己组装要用的垫片去毛刺。团队用 API 迭代了 4 天就拿到这个模型。机器人给自己做零件，有点套娃。

https://t.co/kNGvoJhQ3O

[HawksT @John43497285]

@swyx @aiDotEngineer @SydneyScouts When even the people organizing the speakers could probably give a talk themselves lol

[HawksT @John43497285]

@WSJ Going back to look for it and getting caught on camera is quite the ending.

[HawksT @John43497285]

@Reuters Can a company push back on military use and still be considered a reliable supplier? That's a tricky precedent to set.