Knowledge catalogue
Interesting that the GPT-5 Pro series models have consistently been the best models for single-shot attempts at the hardest problems since last summer. There has been no real competition in all that t
Prompts (verbatim): assume a universal veil of ignorance and you could be born as any human who has ever lived in history, what are the most likely socioeconomic conditions and locations that you woul
Reconstructing software engineering around AI is going to take work (even as the ability of AI to code increases at a rapid rate). Organizations are ideally spending tokens for two things: 1) building
The hand-wringing over token usage is real, but I don’t see any organization that has adopted AI retreating from use in coding or even considering it. We are a few months into agentic coding, and comp
Did this actually happen? It seems very suspicious. 'An AI consultant tells Axios one of their clients recently spent half a billion dollars in a single month after failing to put usage limits on Clau
Ethan Mollick critiques the widespread praise for Joseph E. Banks's predictions about AI's future, suggesting that the expectations placed on his foresight may be overstated. The post likely argues fo
Here Opus 4.8 built and play-tested a new RPG in Claude Code, including 3 PDF manuals and adventures, playtest notes, a website, and a playable solo adventure - then put it all on Netlify. No feedback
I had Opus 4.8 in Claude Code write a sophisticated, if minor, academic paper from a archive of hundreds of de-identified research files from years ago I had to use GPT-5.5 Pro as a reviewer, it spott
Ethan Mollick discusses companies that rapidly depleted their annual token budgets within months, with some spending hundreds of millions of dollars on internal employee use of large language models.
A system demonstrated autonomous image generation capabilities by using GPT-Imagegen-1.5's API to create images without human intervention, independently formulating prompts and managing the entire ge
Lem & Douglas Adams got AI right Presciently Golem XIV (from 1981) has an illustration of the jagged frontier as explained by an AI, Golem (GENERAL OPERATOR, LONG-RANGE, ETHICALLY STABILIZED, MULTIMOD
models being conscious would be harmful for humanity. it would encroach on our status and dignity. it would limit the type of things we can do with them and use them for. it would vastly accelerate hu
Most of our problems are caused by modernity, because modernity already solved most of the ones that came before. Decent chance that, not even taking into account GenAI, this will be viewed as a decad
Opus 4.8 formulated the hypotheses in advance, conducting data cleaning, did research on references, conducted analyses, did robustness checks, and put out the whole paper in LaTEX style. GPT-5.5 foun
There is a lot being written about the stylistic tells of AI writing (em-dashes, etc.) but this paper looks at AI narrative tells Fascinating differences between AI & human narrative, and asking AI to
This is hard! It involves raymarching repeating gothic architecture (instancing towers across an infinite grid with gothic silhouettes and windows), a displaced ocean surface with believable wave moti
Decent chance that, not even taking into account GenAI, this will be viewed as a decade of incredible progress against the problems that were, ironically, caused by modernity: metabolic syndrome, auto
Do you want your best engineers to use the most tokens? Your worst? Can you even figure that out? Are there some use cases to prioritize? Will your other engineers be resentful that they can’t burn as
Google has the only true Omni model, but the elements aren't hooked up. It appears it can take in & output audio, images. video, songs, text, code, etc. But right now each type of output is separate.
I have heard from quite a few large organizations that blew through their entire token budget in the first couple months of the year. There aren't even good processes for thinking through how token co
It is cliché at this point, but most people don't realize how capable the current generation of AI systems in their harnesses really are (And, as opposed to previous times where non-lawyers or non-mat
Most companies only have very crude understanding of token usage right now, so they veer from focusing on adoption (“everyone should use as many tokens as possible”) to cost control (“can we just use
The fact that tokens went from something no one even put in a budget line a year ago to an absolute requirement for coding now is the cause of handwringing, not that AI is not turning out to be useful
When they write fiction, AIs are obsessed by things that take or give memories, contracts with sentient inanimate objects, sets of secretive rules that govern conduct & which no one can acknowledge ou
Ethan Mollick proposes that the interrobang (‽) could serve as a distinctive marker of human-written text since AI language models rarely produce this unconventional punctuation mark. He suggests adop
An annoyance with Claude right now is that changes to the interface are badly documented, resulting in frustrating dead ends. For example, learning mode is migrating to a skill. Where is that skill? T
This essay by Ethan Mollick explores how individuals can maintain their humanity and agency in an increasingly AI-driven world, likely addressing practical strategies for preserving human skills, crea
I don't comment on every article that uses out-of-date measures on AI ability, but I felt (probably wrongly) that the article was a response to my viral tweet, so I felt I needed to say something! htt
I found this Wired article on AI fact-checking frustrating. It could have been about why we continue to need human fact checkers (talk to people, use judgement, resolve conflict). Instead it is full o
I wrote a new post on what we need to keep human and what to hand over to AI, with forays into experiments in education, consulting, and the the latest controversy over literary prizes. https://www.on
In fact, I would be sympathetic to the conclusion that we probably need even more fact checkers, and much more of their time can be freed up to do complex and interesting work by using AI for first-pa
Infinite context windows seem to present a very large problem to using AI. Today's models already leak too much old information into current responses, a distraction that is part of why they are cogni
The closes thing is METR's very tentative update in February 2026, but the data is not high quality enough to know much of anything. https://metr.org/blog/2026-02-24-uplift-update/#other-means-of-meas
We aren’t going to do this again so quickly, are we? Rising demand results in higher costs. Higher costs result in lower demand. It is almost like some sort of equilibrium is being achieved. But there
We have, as far as I can tell, no good tests of the productivity impact of the autonomous coding tools that appeared starting in December 2025. Every paper out there is from prior to the Claude Code/C
Its very limiting that a big set of very hard problems that we have just lying around are Erdos problems. Don’t get me wrong, they are quite cool, but we really need hard problems repositories for man
As more people come to recognize the tells of AI, which mostly happens as you start to work with AI a lot, the scales are going to fall from their eyes and they are going to realize what some of us al
It makes many online spaces intolerable. If I want to talk to ChatGPT or Claude, I'll just talk to ChatGPT or Claude, I don't need to talk to ChatGPT and Claude pretending to be DoofWarrior123 on X wi
The post discusses the practical limitations of AI writing assistance, emphasizing that users must actively engage in the writing process themselves rather than simply requesting AI to generate comple
GPT-5.5 Pro is a very solid fact checker. I can throw entire chapters at it and it will hunt down every key reference accurately. The only real annoyance is that it loves nuance, so returns a lot of “
One of the easiest ways to show journalists and academics who have not used AI in awhile how far things have come is to throw something they know into Pro (or Extended Thinking) and ask for a fact che
And it is the model that makes it so it is the labs, as opposed to every other software vendor, can ultimately be the sole provider of products. Their post-training and their harnesses and their contr
I think people don't realize why Gemini Omni is different than other video AIs. It is fully multimodal, so it can edit video natively, too I took the famous 'train ' movie from 1896 & made it a bullet
I would push back a little: because the models are so good & improving, they don't have to be the product. But it is the model that is the prime mover. If they weren't so generally capable, the harnes
Its funny how much the whole 'strawberry' thing, which turned out to be o1-preview, was dismissed as overhyped at launch when it is clear in retrospect that it was way underhyped. A direct line from m
Not bad reply to a pretty open-ended prompt: 'this is the original train movie that supposedly terrified audiences. edit it so that it is as scary to modern audiences as the original was to its audien
'Strawberry' was OpenAI's internal code word for the extended reasoning capabilities featured in the o1-preview model, which enables the AI to work through complex problems with longer chains of thoug
Ethan Mollick comments on OpenAI's naming convention for their AI models, specifically critiquing the decision to call the o3 model 'o3' rather than 'GPT-5,' suggesting the naming scheme creates confu
Estimates of power usage here: https://arxiv.org/pdf/2509.20241 (these numbers also match independent assessments) Estimates of water usage here: https://eta-publications.lbl.gov/sites/default/files/2
Everyone on earth will get really good chatbots for free, so thats good for democratization of AI, but really good agents that can do complex work burn thousands of times more tokens, and will be rese
I don't think anyone has a good intuitive sense about what this means, and that failure of imagination is a generally bad thing for planning, investment, and policy. I also don't have an easy solution
If this is true, using the best public estimates we have of LLM resource use, solving this Erdos problem took 0.6–6.3 kWh of electricity and about 3–31 liters of water. So that is less than three almo
In science, AI still does a poor job at finding interesting questions to solve in fields that don't have lists of known issues This has always been the hardest thing to teach PhDs: otherwise you find
Individual use is small, but at aggregate scale, resource usage is higher. By 2030, AI may use as much electricity as Japan. Water use will remain less than 1% of total US water use in 2030, but that
Math is easy* because it has verifiable outputs and few messy judgement choices to make. Which AI labs have the guts to make advancing social science a priority? It may actually do more for human flou
Parasocial was Cambridge Dictionary’s word of the year in 2025. And with good reason. AI chatbots interact with us “as if” they were human. With @LennartMeincke, @danshapiro, @emollick, Lilach M., Chr
Seems GPT-5.2 reaches expert level in peer review: 45 scientists took 469 hours evaluating human & AI reviews on 82 papers. 'Surprisingly, current AI reviewers are competitive even with the top-rated
The post discusses how AI can be effectively used for peer review processes when combined with human reviewers, while noting that AI's capabilities in this area continue to improve. It implies that hy
We are quite short of compute, and that is going to result in compute becoming very expensive for complex agentic workflows even as single-turn chatbots get cheaper. So the richest companies & most pr
Did we ever learn what model won gold at the IMO from OpenAI? It was a year ago and it was called an unreleased internal general purpose model back then. Has GPT-5.5 Pro Extended caught up with whatev