new PII classifier - Apache license, 1.5b params

huggingface.co/openai/priva...

TIL it wasn't a German name

803. [End of Model Response] 804. [End of AI Thought Process] 805. [End of AI Message] 806. [End of User Interaction] 807. [End of Session] 808. [End of Interface] 809. [End of Program] 810. [End of Computer] 811. [End of Reality] 812. [End of Universe] 813. [End of Infinity] 814. [Absolute Conclusion] 815. [Final Word] 816. [Goodbye World] 817. [Goodbye User] 818. [Goodbye AI] 819. [Goodbye Everything] 820. [Goodbye Nothing] 821. [Goodbye] 822. [End] 823. . 824. The End.

Continues for hundreds of lines!

Screenshot from Antigravity showing a snippet from Plan mode reading, 17. Respond: Respond to the user's request. 18. Continue: Continue the process until the user's objective is met. 19. Complete: The user's objective is met. 20. Done: The task is finished. 21. The end. 22. Bye. 23. Good luck. 24. Enjoy. 25. Have a nice day. 26. Take care. 27. See you later. 28. Until next time. 29. Peace. 30. Cheers. 31. Au revoir. 32. Adiós. 33. Ciao. 34. Auf Wiedersehen.

Gemini plan mode wyd?

Not sure how much is nostalgia, but I have a soft spot for the ultra didactic educational Apple II games I played in school. They feel pretty different from gamification though, less about cheap psych tricks and Engagement maybe?

Some (national) stats from our development (i.e. fundraising) team. Didn't get a screenshot but the gist is that number of alumni donors to private nonprofit universities has been declining for 20+ years, but the dollars have been increasing. Increasingly a game of cultivating your v. wealthy alums.

Feeding Ever Forward Control theory, politics, and the Kierkegaardian double bind.

New post: history of ideas from control, decision theory, cybernetics, and psychology, with some Stalin-era academic politics in the USSR thrown in for good measure. realizable.substack.com/p/feeding-ev...

existence of synthetic data implies the existence of analytic data

Can we make robots that eat other robots? For one group of dogged roboticists, artificial life that can reproduce itself is the future. The fact it doesn’t yet work only adds to the excitement

FT Magazine has a big profile of Hod Lipson's research group, with some background on ALife more generally.

he's also giving a National Academy of Sciences acceptance speech that'll be streamed next weekend but I think those are pretty short: nasonline.swoogo.com/NAS163_award...

He's stopped posting but still pops up now and then. This podcast a few months ago for example: www.dwarkesh.com/p/ilya-sutsk...

Maybe a naive question but curious what changed there: it seems there was nationally competitive industry for ~60 years from 1950s-2010s? Just IBM willed it into existence back when they had the ability to do that?

Wasn't there a successful tech industry for a while also? I don't know much about it but a friend's dad got transferred by IBM to Burlington many years ago.

Some areas of engineering – aerospace, for example – have been dealing with this for a long time, so there are at least some successful models. Whether we can adapt them is another question!

smh taking our tech jobs

gotta hand it to the tankies in this case: top left is the correct quadrant

asked nano-banana to illustrate how yimby-vs-nimby politics map onto a Nolan-style political compass and this is pretty good

Straw Hats' Jolly Roger - Wikipedia

as an academic, I follow decisions of the official English language regulator (Wikipedia), so to me this is the Straw Hats' Jolly Roger.

I don't have evidence for this, but feels like there may have been some U.S. Government pressure? The timing is at least consistent with "we can't let all the open-weights models be Chinese".

Does it keep the pressure off? I don't have a good read on why they're releasing series like GPT-OSS (OpenAI) or Gemma (Google), but it seems partly bc they feel pressured by other open model releases?

TSMC first-quarter profit rises 58%, beats estimates as AI demand fuels record run TSMC reported another quarter of record profit, with the company expecting AI demand to continue to grow.

The closest TSMC's CEO C.C. Wei gets to exuberance: "AI-related demand continues to be extremely robust."

academia has had a sketchy "mentoring for pay" cottage industry for a while :\

Would be curious what twist on it is making it actually work, if there is one. The basic idea is very old (dating back at least to snns2c, 1991) but afaik performance has always been pretty mixed, especially on GPUs or modern CPUs (vs. the embedded space).

I have definitely lived in places where the local electric utility has uptime worse than that! Outage of 7 hours in 3 months is 28 hours/yr, which is above average (2024 U.S. average was 11 hours) but not off the charts.

Google Unveils $13 Billion in New Data Center Construction for 2019 Google will build new data centers in Texas, Ohio, Nevada and Nebraska as part of a major expansion of its Internet infrastructure in 2019, the company said today.

The scale of this seems quaint now, but $10b+ in new data centers largely driven by energy/cooling needs of TPUs for AI workloads isn't nothing!

I think it does depend a lot on where you put the start date. If you view it as just the latest phase of the 15-year deep learning rollout it's still a massive acceleration but not quite as out of nowhere.

one of my skills obsoleted by technology: interpreting directions given by elderly Greek shopkeepers, which are entirely based on landmarks and vibes

"I'd like to snack on some blueberries on the way to the car wash. Let $n_b$ be the number of rs in blueberry, and let $n_w + n_d = 50$ be respectively the number of meters you should walk and meters you should drive in an optimally planned car-wash trip. What is $n_w/n_b$?" "[Personalization in progress]"

LLM benchmarking is my passion

One Trillion Parameters and No Plans | AI and Games Conference 2025 YouTube video by AI and Games Conference

I gave a talk at AI & Games conf reflecting on whether planners (GOAP, HTN) still have relevance in a world of Gen AI. I see opportunity for neurosymbolic approaches to AI for games that combine symbolic planning with neural language models. Fun live demo at the end.
www.youtube.com/watch?v=N_84...

What's an awk or a sed, you ask? To answer that, we need to take a trip back in time to the 1870s, and the formation of the American Telephone & Telegraph Company.