Google's LLMs are cracking me up today. I fed some text into Gemini 2.5 Pro that contained a misspelling. Gemini guessed that what the writer meant by "veriority" was probably "verisimilitude". Gemini has so much faith in human intelligence. Yes Gemini, the person who can't spell "variety" definitely actually meant "verisimilitude" and also definitely knows what that means.
Gemma 3 seems a little too fine-tuned to be agentic, and kept hallucinating that it was checking web results even when it had no tools available. The amount of grovelling that Google RL'd it on is really quite impressive. It was hitting every part of an apology: acknowledging its precise mistake, self-flagellating, and promising to do better in the future.
Jonathan
_@jonathansm.com
npub1uqee...jckg
Hacker, cypherpunk. All memes are my own.
Notes (18)
I’ve been using Aider recently for AI pair programming. It seems like the best setup I’ve seen so far for managing context and doing real programming work.
It is impressive tho how insanely expensive Anthropic’s models are. Google has the same or better performance for just a fraction of the price.
https://www.bbc.com/news/live/c9wpq8xrvd9t
Wait what? In both Spain and Portugal the entire country lost power?
https://www.cato.org/policy-analysis/illegal-immigrant-incarceration-rates-2010-2023
New Cato institute report shows that both legal and illegal immigrants commit less crime than native born Americans. Even if you exclude all native born black people, immigrants *still* commit less crime. Can Americans all finally shut up about immigrant crime now?
https://simonwillison.net/2025/Apr/26/o3-photo-locations/
Fascinating article from Simon Willison talking about how good o3 is at guessing your location from a photograph. We’re only beginning to learn the new capabilities of reasoning models using tools during the thinking process.
https://www.startupdaily.net/topic/artificial-intelligence-machine-learning/eye-thy-with-my-ai-australian-radio-network-digitises-diversity-with-an-artificially-generated-asian-female-presenter/
Wow. The company behind Gold and iHeart radio wanted more diversity so instead of hiring someone they made an AI radio host that sounded like an Asian woman using TTS to run hours of music per week without telling anyone.
https://theappendix.net/posts/2014/02/darwins-children-drew-vegetable-battles-on-the-origin-of-species
https://en.wikipedia.org/wiki/Onfim
I love when we find kid’s drawings from centuries ago.
https://en.wikipedia.org/wiki/Trump_Tower_(novel)
Donald Trump wrote a bad romance novel? Why did he think this was a good idea?
https://alisoncrosthwait.substack.com/p/a-ketamine-addicts-perspective-on
https://x.com/powerfultakes/status/1892003738929238408
New double prong theory for Elon’s mental decline. Perhaps it’s a combination of both ketamine and genetic predisposition. For some reason there seem to be a lot of incredibly smart people who took ketamine and fried themselves. Felix Hill is an example here.
Probably to make his bid for Pope. He already had the look.


https://velvetshark.com/ai-company-logos-that-look-like-buttholes
Dear god now I can’t unsee it.
You don’t actually know the Ten Commandments. The ones you just thought of were smashed by Moses and then God replaced them with a new set of Ten Commandments. The new Ten are very different from the original Ten. Go read Exodus 34.
Weird realization for the day.
https://www.noahpinion.blog/p/will-this-be-the-chinese-century
Bearish article on the future of America. Interesting insight that when information in the internet age spreads nearly instantaneously, being innovative doesn’t give you much of an advantage. Making incremental improvements to bring that innovation to market and actually implement it is what leads to world dominance.
One point of disagreement though. I think everyone people were underestimating Trump but have swung way too far to overestimating his impact in 4 years. The Democrats and the bureaucracy may have been initially thrown off balance but they’ll recover and retard Trump’s initiative.
https://www.molsonhart.com/blog/america-underestimates-the-difficulty-of-bringing-manufacturing-back
Lots of good points here. One thing I hadn’t considered is that tariffs make it really difficult to establish manufacturing here. The cheaper option may simply be to do as little manufacturing in the US as possible and only import the finished product rather than deal with the complexity of an industrial supply chain and trying to import the intermediate products to do the manufacturing domestically.
https://www.wsj.com/politics/elon-musk-children-mothers-ashley-st-clair-grimes-dc7ba05c
I guess… it’s good Musks actions match his rhetoric on repopulating the planet? Besides that this seems like super weird fetish behavior.
https://ai-2027.com/
https://openai.com/index/introducing-o3-and-o4-mini/
We appear to be hitting the "Reliable agent" section of the AI takeoff timeline. o4 is probably a real human level coder at this point if the Codeforces benchmarks are anything to go by.
