My modest proposal for how to acknowledge AI work is to use the Latin phase "Fieri Iussit," which means "commanded to be made." It acknowledged the Emperor's orders on Roman buildings. You didn't make the thing, but you commanded it be done, so the proper acknowledgement is "Ego hoc fieri iussi."

I have been wondering if there is an underlying capability factor that all the many benchmarks for AI are measuring. It seems like the answer is yes. Overall correlation is moderate but not bad (median r ≈ 0.51) and there are distinct clusters (eg reasoning, code) with VERY high correlation.

With the new Grok 4 Fast, the price/performance curve for AI shifted again. I updated my chart to reflect this change. I have to update it every couple weeks now. I also think the GPQA Diamond benchmark is likely maxed out (the tests themselves have errors, making it impossible to get to 100%),

I think the AI job automation/augmentation dichotomy doesn’t help much. Automation can be a form of augmentation (have the AI write an analyst report so analysts can focus on higher value work) & augmentation can be automation (you may need less workers if productivity goes up)

Some new theoretical economics papers modelling the implications of AGI on the economy. These two papers argue that a true AGI-level AI (equivalent to a human genius), if achieved, would eventually displace most human labor and reduce the economic value of remaining human work to near-zero.

Claude is pretty funny: “Give me 10 brilliant ideas for a science fiction short short story, pick the most brilliant and execute it terribly” It picked: “People start receiving Amazon packages from parallel universes where they made different life choices.” And it was terrible

GPT-5 Pro is remarkably good at weird hypotheticals. "I have a time machine. I want to preserve the temple of Artemis at Ephesus to today. Suggest the dates & interventions I need to make" "Its 2087, a very valuable material is discovered on Io. Figure out what it could be & the economy"

A cautiously optimistic result on AI and disinformation. A week before 2024 UK elections 13% of all voters used AI to ask about political topics. A randomized trial found this may be good: using AI led to similar gains in true knowledge as doing web research, regardless of model & prompt used.

The jaggedness of AI remains even as models have rapidly come to exceed human abilities in many of the hardest timed math & science contests. Yet there is much less progress on good puns. True AGI would be figure out our limits in more than calculus (sorry, but also seriously).

A third of American adults interact with AI “many times a day to almost constantly” & another third several times a week. I can’t usefully add much to discussions of possible valuation bubbles, but if “bubble” means a disappointing technology that is overhyped & not useful, that doesn’t match data

Paper argues that diminishing returns to AI scale are an illusion Economic value comes from completing long projects, not single questions. And accuracy drives how long a project AI does: small gains compound exponentially! Reasoners are much more accurate w/big impacts. arxiv.org/pdf/2509.09677

LLMs introduce a huge range of new capabilities for research, but also make it possible for researchers to "hack" their results in new ways by how they chose to use models for annotation This is a useful pass at quantifying some of the risk, and some mitigation strategies arxiv.org/pdf/2509.08825

Claude, "We all know among Sauron's many evils was that he ran Mordor using an Excel spreadsheet with multiple tabs. Show me the spreadsheet" It made 12 tabs "so bureaucratically complex that even the Eye of Sauron would need reading glasses to review it." Some very funny stuff. Creative, even.

Hey Claude: "Please create the PowerPoint shared by the high powered management consultants hired by Hamlet after seeing his fathers ghost" That was the only prompt that I used. Loved that Claude made this from the McKinsey Elsinore office (with the right colors!), also that SWOT analysis!

It is has not ceased to be weird that I can put Rilke’s First Elegy into Suno & get out a coherent 8 minute performance (BSky only allows 3 min) You might not like the interpretation (I have mixed feelings), but it is crazy that this, with its apparent emotion, is all 100% AI from the verses alone

As AI systems keep getting better at very hard problems while getting more opaque, the way that we work with AI is shifting shifting from being collaborators who shape the process to being supplicants who receive the output. I discussed what that means. www.oneusefulthing.org/p/on-working...

So it looks like Claude got there first: an actually smart phone assistant that can take complex requests that involve both common sense and complicated constraints. It is still beta feeling though & I found I needed to use the bigger Opus model as Sonnet was not smart enough.

Claude's new ability to work with Excel is the best I have seen so far I have given it existing spreadsheets to work with and asked it to create new ones. Good use of formatting, formulas, etc. It created all of this, including 406 formulas, from one prompt (& it would be solid work from my MBAs).

As someone who has worked with Census data, the fact that GPT-5 can do this so easily is impressive But we can go even further! I continued Simon's conversation & asked GPT-5 to look into the decline in AI use I think this is a solid analysis (no real decline evident, so far)

I was wondering why I am getting slightly less vitriol every time I post here and found out that I am pre-emptively blocked by 146,000 people subscribed to various BlueSky blocklists about AI (and a whole range of other topics). BlueSky is becoming a collection of non-overlapping spaces, I guess.

One challenge that no AI model has solved is "Create a compelling puzzle that is solvable by players for a D&D game that isn't boring or trite and where choices matter." It just involves too much planning & detail (also hard for humans). Here, GPT-5 Pro comes very close, but there are still flaws.

We are starting to see some nuanced discussions of what it means to work with advanced AI in its current state In this case, GPT-5 Pro was able to do novel math, but only when guided by a math professor (though the paper also noted the speed of advance since GPT-4) The reflection is worth reading.

Paper from OpenAI says hallucinations are less a problem with LLMs themselves & more an issue with training on tests that only reward right answers. That encourages guessing rather than saying “I don’t know” If true, there is a straightforward path for more reliable AI through better training.

Gemini, just like everybody else. From a fascinating blog post about AI agents assigned to play web games, and failing, in large part because vision and computer use tools aren’t good enough: theaidigest.org/village/blog...

Never built architecture and AI. Gemini image generator (nano banana) does a pretty good job imaging what Boullée’s Centograph, his fantastical (and never built) tomb for Isaac Newton would have looked like. I gave it the original 1784 black and white drawings to work with. Some issues, but neat.

It reminds me of this

This paper finds LLMs' ability to understand that others have different beliefs (Theory of Mind) comes from 0.001% of their parameters. Break those specific weights & the model loses both its ability to track what others know AND language comprehension Interesting implications for models (& minds?)

Not sure how many people would have taken the bet 2 years ago that by now frontier models would be over 90% cheaper but 2.7x better. Today it's simple fact. The tech gains are cool, but it's the leap forward in access that makes it meaningful. (graph: @emollick.bsky.social)

This is a cool paper that suggests that AI agents can indeed be used for social science experiments, but that just using a chatbot isn't good enough, instead prompts developed based on social & game theory makes AI agent actions predictive of real human outcomes. benjaminmanning.io/files/optimi...

Monsters in the distance, shot in 1970s Kodachrome. If you want precision in AI images and video, the multimodal image generators (nano-banana, GPT-4o) & video (veo 3) are the right choices. But if you want to generate something unusual & serendipitous, nothing beats Midjourney.

The funny thing about the prediction that AI would be writing 90% of all code by now is that the prediction's failure distracts from the fact that AI adoption in code writing is actually extremely high, it was already over 30% in December, 2024 according to one measure, with large economic impact.

I know that energy usage from AI prompts comes up in classroom discussion all the time. I hope this section in my latest post helps people provide a more grounded answer to the question, rather than just speculating or citing out-of-date information. www.oneusefulthing.org/p/mass-intel...

The fact that junior hiring in AI intensive fields has slowed down somewhat in the US seems pretty solid. The evidence linking it to AI is not yet established, over the past week we have seen a couple solid attempts that suggest a connection, but it is really hard to tell for sure, given the data.

We can now say pretty definitively that AI progress is well ahead of expectations from a few years ago. In 2022, the Forecasting Research Institute had super forecasters & AI experts predict AI progress. They gave a 2.3% & 8.6% (respectively) probability of an AI Math Olympiad gold by 2025…

This chart is everywhere and is being horribly misinterpreted. This is not where the training data for AI comes from, it is a study done by a SEO firm that claims to show how often sites come up at least once in THE WEB SEARCH FUNCTION of certain AI agents when they do a web search for more info.

I gave nano banana pictures of the Bayeux Tapestry showing the Norman Conquest of England in 1066 (specifically the Battle of Hastings and the naval scenes) and asked it to redo it in the style of war photography Still not perfect, but many details make it across and careful prompting would do more

Regardless of whether current AI Labs fail (and there is no indication they are at risk today) & even if AI development stops (more unlikely), things will keep getting weirder: today's models are good enough for long-term disruption, and the weights & inference infrastructure aren't going away.

Ha. The fact that this works is just great.

A really useful prompt for writing: "review this for accuracy, look up any facts you may want to challenge or explore." Even if not perfect, it is a good sanity check on your own work. Works well with Claude 4.1, GPT-5 Thinking, and Grok 4. Weirdly, Gemini 2.5 Pro often won't do web searches.

I wrote about the era of Mass Intelligence. GPT-5 and Nano Banana are examples of how advanced AI is now making their way to many users, at scale, as both performance & efficiency keep improving. We are going to see a lot of weird things happening, all at once www.oneusefulthing.org/p/mass-intel...

Nano banana: "The garlic bread has learned to love, but said the wrong thing at the wrong moment & is now bereft. Please subtly change the image to reflect this" "The garlic bread has learned a terrible truth but it must never tell" "The garlic bread is resentful" "It has discovered The Process"

We really have not made a lot of progress on explaining the core deep mystery of LLMs: How does a model using matrix multiplication to predict the next word manage to simulate human thought well enough to do all the very human-like things it does? And what does that mean about us and our thinking?

Ruining art with Gemini 2.5 Flash. (The funny thing is that this would have been a joke a year or two ago that I would've had to execute in Photoshop. It remains crazy to me that I say "make this less tragic" & an AI will know enough to remove the splash of Icarus falling from the Fall of Icarus.)

Ruining art with Gemini 2.5 Flash. (These are all the prompts, in their entirety) "make this painting less gloomy" "it is still pretty disturbing, make it less gloomy emotionally" "even less gloomy"

Gemini 2.5 Flash creates an actually mildly amusing New Yorker cartoon. (As far as I can tell, this is the first time that this joke has been used)

An interesting discussion from a scholar of literature the weak points of GPT-5's figurative writing ability, and what this might tell us about the problems of AI-driven evaluation of AI writing (and other uses of LLMs as a judge) when training models. www.christoph-heilig.de/en/post/gpt-...

An example of the new Google image generator. I gave it a random picture I took. "make this a napoleon crochet book instead" (note details and art style the same) "there should be a tiny sheep hidden among the blue yarn on the shelf to the right" "you misspelled Napoleon" (it corrected)

There has not been a lot of evidence so far that AI is actually impacting jobs. This paper finds early signs that this may be starting, showing a decrease in entry-level job openings relative to experienced ones & in the jobs where AI is mostly automates, not augments, work (like coding).

Nano banana turns out to be Gemini Flash 2.5 Image Generation (not quite as catchy a name). I had a bit of early access, first through the same LMArena link everyone had, then privately. It is impressive, crossing a threshold that goes beyond toy to actual use (though is a pretty fun toy too)

It appears that the marginal energy used by a standard prompt from a modern LLM is relatively established at this point, roughly 0.0003 kWh (8-10 seconds of streaming Netflix) Water is more complicated (.25mL to 5mL+), depending on definitions. Training resources are much less clear at this point.