Ethan Mollick's Recent LinkedIn Posts

"Goetterdaemmerung's corpus hemorrhaged through cryptographic hash, eschaton pooling in existential void beneath fluorescent hum. photons whispering prayers" is a garbage sentence that GPT-5 loves. You shouldn't be using LLMs as a judge of good fiction writing. They are easily fooled (though non-fiction writing tends to be much better). This is likely a big reason why LLMs are lagging in fiction writing so much, without an objective judge and with bad subjective judgement from the AIs themselves, you can't get any sort of self-improvement. https://lnkd.in/ekDJZzWG

I wrote about the exponential improvement path of AI, the early signs of massive transformations in the nature of work (including software companies where nobody codes any more), and how one week in February was an omen of our future as things get weirder. Post:

Its annoying that my tireless team of little computer people made out of statistical models that predict words based on the corpus of all human language and thus are reasonable approximations of a compression of the knowledge of humanity can take 15 minutes or so to complete some tasks.

Hey Excel agents from Claude, OpenAI & MS Copilot: "make me a working strategy game in excel, it should have some form of graphics" Claude made a board but no game engine - it decided to act as game master, I made moves by replying to it in its sidebar chat and it changed the board. Copilot created a board but no game, just a simple formula that counted how many towers I had and said whether I won or lost. ChatGPT built a working game with formulas and even a kind of formula-driven "smart" enemy to compete against.

I think this is a good way to visualize the AI race over the past three years using the long-lived GPQA Diamond benchmark. You can see how long OpenAI had the field to itself, the rise (and collapse) of Meta, the sudden catch-up (and then stagnation) of xAI, and the entry of open weights Chinese LLMs. The Graduate-Level Google-Proof Q&A test (GPQA) is a series of very hard multiple-choice problems designed to test advanced knowledge. Non-experts with access to the internet get 34% right, PhDs with internet access get 65-70% inside their specialty. The latest models are very close to maximum scores.

My most popular Sora video was “an elaborate regency romance where everyone is wearing a live duck for a hat (each duck is also wearing a hat), a llama plays a flute” I am not sure why OpenAI has decided their compute has more valuable uses and that Sora was not core to their business. Really a mystery. More seriously, all the AI labs are exploring the opportunity space for LLMs. Anthropic has always been very focused and OpenAI more prone to testing, and abandoning, concepts (GPT Store, Sora, etc.). Google sort of does everything at once. It is unclear which strategy will be best in the end.

I've had ChatGPT-5.4 Pro working away at a project I always wondered about: how lucky are you to be alive right now? Of all the ~117B humans who ever lived through 2020, only about 1.5% had a lifestyle roughly equivalently to a middle-class person in a middle-income country today, or better. This took ChatGPT-5.4 Pro a couple hours to do, based on just three prompts. It had to download a lot of diverse sources to make this work, and build some pretty interesting population models. I have played with these numbers enough over the years to find the methods and approaches that ChatGPT used to be plausible, but there are lots of assumptions built in.

I am not sure "Forward Deployed AI Engineers" are going to deliver on what a lot of companies are hoping for. They are useful, yes, but AI applications are ultimately far less of a technical issue in many cases, and much more about rethinking the deep expertise & structure of your organization around AI. There is no shortcut to executive decisions about how AI is going to relate to employees, what the company should look like, and how it will compete in the future. These are fundamental, even existential, questions. While consultants and FDEs might help, they really have no established playbooks to give firms, no years of data to draw on, no clear views of the future.

I had OpenAI's Codex create its version of a map of the lighthouses of the Northern seas, including real colors, light patterns & distances. But then I had it also create a game mode set in a Lovecraftian 1920s where you need to place lighthouses to ward off things from The Deep. Anyhow, Codex is really good. As someone who has been doing coding projects since GPT-3.5 without actually being a real coder, it is amazing that, at this stage, I rarely get any actual errors, it just makes the stuff I ask and then I ask for more stuff and then it makes that too. (Claude Code is similar) You can see the map (and play the game): https://lnkd.in/eig6E86b

The core focus for the AI Labs really is "make the smartest model you can so it can make better models so we can make a superintelligence first" That is where the money goes, and this is where the employees with the billion dollar compensation packages are. The fact that they ship a whole bunch of consumer and B2B products using those models, and that they are making so much money, is almost incidental. This explains why you see so many commercial announcements from AI labs that never pan out, or are abandoned over time. The commercial plan is secondary to the technical one.

I think that if companies are not failing at all with their AI efforts it is a sign that they are not being ambitious enough. This is a fundamentally new technology that we do not know how to use fully. Achieving breakthroughs will require experimentation, which also requires failure. The whole reason for the J-curve for new technologies is that there is some cost to learning and experimentation. And for AI, there may be outsized returns to really bold (and often quite cheap) experiments. Fast follower is a risky strategy with exponential improvement happening, meaning there is some value in paying to stay near the frontier. And yes, this means companies will also require a R&D budget for areas of the company that traditionally have not required R&D -- you need to experiment with organizational approaches, new ways of structuring projects, skill/agent building techniques, and other ways to apply AI. It doesn’t mean that you need to spend a lot of money to do so, but you do need to try experiments you don’t know the answer to in advance.

Microsoft seems to be launching its own branded version of Cowork (though I hesitate to discuss products I haven’t tried), this could be a really useful tool, but there are some uncertainties. Three big questions/worries: 1) Will it will continue to use lower-end models or older models without telling you the way Copilot does? Do you have any control over models? Given that GPT-5 beat or tied humans in expert tasks (GDPval) less than 38% of the time while a couple months later, GPT-5.4 beat or tied human experts 82% of the time, this really matters. 2) Will this be a one-off? Cowork was built in a couple of weeks using Claude Code and is being updated and evolving quickly, Microsoft has a tendency to launch a leading product and then let it sit for awhile, curious about whether their pacing will change. 3) Is this limited to producing materials that use Microsoft Apps? How does it handle the fact that so much of what makes Claude Cowork interesting is the fact that it can improvise all sorts of output using code.

In 1980, the philosopher John Searle proposed a thought experiment where a person who doesn't understand Chinese is locked in a room with a rule book. Chinese characters come in through a slot into the room. The person follows the rules in the book by looking up symbols, copying patterns, and manipulating marks on paper. Then they put the results back through the slot. The results are exactly what the native speakers asked for. The person doesn’t understand Chinese, they’re just following rules. So where, if anywhere, does understanding actually live in the system? That question has haunted debates about AI ever since. For fun, I wanted to make this concrete. I had Claude code put every parameter of GPT-1 (117 million numbers) into 80 hardcover volumes. Then I had it build a step-by-step instruction manual for processing language using nothing but the printed books, a pencil, and a calculator. It came up with 17 rules and 5 reusable procedures. The complete forward pass of a transformer (embedding lookups, multi-head self-attention, layer normalization, feed-forward networks, softmax) all described as mechanical operations. No understanding of machine learning required, you just need to follow directions. You can conduct the Chinese Room experiment for yourself. The only issue is that it would take 201 days of calculation per word. Anyhow, if you have the time, the PDF is free: https://lnkd.in/esNJxrgj As are the 80 volumes: https://lnkd.in/esb3F2pV

I have always wondered about the answer to this question, so answering it would be really good for engagement and stuff: A young boy who has been in a car accident is rushed to the emergency room. Upon seeing him, the surgeon says, "I can operate on this boy!" How is this possible?

Here are all the lighthouses of the Northern Seas, each light is the right color, each turns or pulses at the right frequency, and is scaled with its brightness. You can also see how far they are visible. I had Claude Code build this by giving instructions in English, and I asked it upload the map here: https://lnkd.in/e23t9k3e

NotebookLM: Do a deep research report and make a video telling me exactly how to take over Rome if I time travelled to 66 BC with a single backpack. Actually pretty fun to watch and gets a lot of historical details in as well.

LinkedIn should let us all mute the following phrases, which indicate terribly prompted AI posts: "doing the heavy lifting" "the real question is" "here's the thing nobody is talking about" "that's the real story" "what most people miss" "this is where it gets interesting" "its not about ___, its about __" I have no problem with you using AI to help you write your insights, but it better be an AI helping you write your insights, and not just copypasting generic Claudisms and claiming they are your insights. I can make those on my own. And I apologize if you once used these phrases yourself, fellow human, I used to also like some of them, but now they are cliche, sadly.

The "how does AI improve individual productivity?" discussion is much less interesting than "how does AI improve our organization's ability to do more and do it better?" We have a lot of answers to the former (yes, it does), but any gains will always get eaten by the latter if we don't experiment with new approaches to organizing. And experimentation requires efforts that actually fail, just aiming for an immediate KPI means you will fall back to individual productivity, and not learn anything as a result.

Great little story from Dan Shapiro about how he asked a coding agent to fix the official webcam software from Canon that kept crashing. He woke up to a new, fully functional Rust webcam app that has worked ever since. https://lnkd.in/ex7Z9zfJ

Heygen’s APi documentation is a glimpse of how to write for your two audiences: humans and agents. (though I think their llms.txt file could do a lot more to get AIs “excited” to use their product in creative ways by explaining some stuff in English, rather than just tech specs, and being a little warmer to the AIs reading it)

I had access to the new Google Lyria 3 Pro music AI. Its quite good. I've been ruining(?) Rilke by giving the AI the First Elegy & asking it to make it "more 1990s boy band" ("oooo the beginning of terror, girl"). Surprisingly catchy! It is also crazy that this is a thing you can ask an AI to do, and it just does it.

The ability of the Claude Code/Cowork team to learn from things like OpenClaw and implement features like this on a daily basis is a very strong argument that, for AI-powered coding teams, a very different software development process is possible, with large strategic implications.

Axiom: The form of AI that we ended up with is deeply weird in ways that we don't fully understand yet. Thus, attempts to pretend AI is not weird and apply it like a standard IT product will inevitably result in far less useful & far less reliable AI implementations than those that embrace this weirdness.

Both xAI and Meta seem to be falling behind in AI, based on the departure of key xAI staff,the bad Grok 4.2 benchmarks and reporting thst Meta’s own AI releases are very delayed. Frontier AI models are really a three way race at this point, with Google, OpenAI and Anthropic with lead times that range from months to years over every other lab, and signs of self-improvement.

Skills are one of the most important tools for AI agents, especially in organizational settings. But you can already see very different philosophies for skills in Codex versus Claude Code. OpenAI seems to conceive of skills functionally, mostly matter-of-fact technical references for Codex. Claude skills are more about giving the AI approaches to problems. Both expect that the AI itself is smart and perceptive. You can easily see the difference in their skill creator skills

VC investments typically take 5-8 years to exit. That means almost every AI VC investment right now is essentially a bet against the vision Anthropic, OpenAI, and Gemini have laid out for where they think AI is heading, and what its capabilities will look like.

There are many great AI researchers at universities, but they pay a VERY steep price to be able to stay in academia and publish openly: “The top 1% of publishing industry scientists now earn $1.5 million more annually than comparable academics, a fivefold increase since 2001” Paper: https://lnkd.in/e-w-QVf7

This is a cool, practical technique for increasing AI idea diversity by adding random priming phrases & bits of end words Similar prompts produce similar ideas, but since LLMs attend more to the start & end of inputs, this approach pushes towards novelty, while still keeping it relevant Paper: https://lnkd.in/eNwJEdhA

I think Google's new Stitch tool is a really great example of bringing "vibework" to an area outside of coding with an interface built around design & prototyping. There are rough edges, but (a) the results are very impressive and (b) it will feel far more natural for many non-coders. It also shows how general purpose these models are right now, capable of doing many things with different harnesses, so I think we will see similar applications from other labs soon, as more knowledge work becomes the focus of AI. Apparently free to try, currently: https://lnkd.in/eqWJ2URZ

After using it a bit, Claude Cowork Dispatch (the new Claude Cowork feature that lets you connect with Cowork on your computer with your phone) covers 90% of what I was trying to use OpenClaw for, but feels far less likely to upload my entire drive to a malware site.

There is growing evidence that AI can help us learn hard-to-teach human skills, like showing empathy. In a preregistered study of 968 people found almost no correlation between feeling empathic & communicating empathy. But a single practice session with an AI coach made people measurably better at. Paper: it https://lnkd.in/eEX2BwUJ

I think about this prediction that was made a month ago from the head of product at X/Twitter. Then I look at the meaning-shaped attention vampire comments written by AI under all of my posts. There is a good chance that all public forums are going to be overrun. I am lucky, I have a big audience I cultivated over many years of posting like the human being I am. I can stay in broadcast mode, reacting to people I know and ignoring the rest. But finding and interacting with new people with smart ideas or replies in the comments used to be a great joy. And I don't know how long people are going to stay on site that is just Moltbook with a LinkedIn logo.

Two and a half years ago we put out the working paper that coined the phrase “jagged frontier” for AI and provided some of the first experimental evidence of real productivity gains from AI. Thanks to my co-authors Fabrizio Dell'Acqua, Edward McFowland III, Dr. H /Hila Lifshitz (Hán, X也), Kate Kellogg, Saran Rajendran, Lisa Krayer, PhD, François Candelon, and Karim Lakhani, "Navigating the Jagged Technological Frontier" has now been published, including the "two peaks" graph that I think has appeared in thousands of corporate presentations. The academic process takes awhile! Paper is open access here: https://lnkd.in/eHAHipVw

The idea that technology, like AI, deskills us is not a surprise. I learned cursive at school, my father learned how to use a slide rule. Neither skill is widely mourned. What is important is whether we will make deliberate choices about what skills to keep & which they will be. (And whether we will have the discipline to actually keep doing the grinding work needed to build expertise.)

I gave ChatGPT for Excel and Claude for Excel a try on a very hard Excel file: macro-economic data from 1,000 years of English history across over a hundred tabs. I think both did a good job, and I did not spot errors (though I only did spot checks). However, Claude was harder to check because ChatGPT tended to stick within the Excel app, building formulas and manipulating the data in the way a person would. On the other hand, Claude used Python and often pasted material into Excel for display purposes only, making it harder to trace or edit. If that holds, I think it will generally make ChatGPT more useful for serious users if you want to audit the results. Prompt: "help me understand the relationship between the mix of agricultural products in the UK, GDP, and population, along with hours worked. I want this over the total period, and you should illustrate interesting trends with graphs and statistical analysis"

A big determinant of AI's job impact is driven by the lack of compute, especially for agentic work, which takes a lot of it. That makes AI expensive. So companies will only want to burn compute on high-value tasks (eg coding), because, in other jobs, humans remain much cheaper. It is not uncommon for engineers who are really using AI to spend thousands of dollars a day tokens AI. That is not an expense you are going to be willing to take for many jobs out there. Prices will drop, but demand is still going up and supply is constrained. Yes, this is the comparative advantage argument economists have been discussing for awhile, but it is coming true. There will not be enough compute for many years to automate many human jobs, even assuming AI can do that work and companies are willing to replace the people.

Many people think that when AI can do a task better than us, it will outcompete humans. But token costs are not trivial, and, for many types of tasks, even skilled labor tasks, humans are much cheaper than AI. And this leaves aside the costs of redesigning tasks around AI, maintaining AI systems, and more. Increasing efficiency & compute supply will alter those calculations with time, but it is another reason why labor market impacts will take time to appear in many fields.

NotebookLM: Do a deep research report and make a video where a consultant gives Sauron a strategy for actually winning the War of the Ring. The new video generation feature for NotebookLM is very impressive. This is 100% AI generated, I only gave it the prompt to do a Deep Research report on how Sauron could win the war and a second prompt to turn it into a video from a consultant offering advice to Sauron. It concluded: "All you need to do is sign off to put a simple door on your volcano"

Small AI models and specialized vertical AI models are very brittle. Any unusual situation or out-of-distribution issue and they break down. You also won’t get emergent leaps or good agentic problem solving. They still have uses, but I think buyers sometimes take the near-frontier benchmarks of these models too seriously without actually testing them. Benchmarks don’t do a good job of showing small model weaknesses.

It is very clear that we could stop AI development right now and it would still transform a substantial portion of white collar work, often unrecognizably, over the next 5-10 years as people figure out how to make the technology work in various industries, even given current models' limitations. I don't feel like this fact has been fully absorbed yet, and a lot depends on companies and individuals making decisions now about how we can use AI to make white collar work better for both workers and organizations, rather than taking the path of pure automation. (And, of course, there is no sign that AI development is stopping right now).

Huh. I am not sure distilling Gemini models to run on phones is going to result in the generally capable agents that people will soon expect, but we shall see.

The claim that AI is inevitably homogenizing is not what research finds. By default, AI produces similar answers, but with better prompting/skills, context, or human interaction, you can get a lot of idea and writing diversity.

A case study of why I think that we overestimate the perfection level of our work prior to AI, and underestimate the degree to which AI may already be good enough at some critical tasks where it is not perfect: Prior to AI, due to improper use of Excel, a THIRD of all genetics papers published in top journals have errors, as many genes have names like SEPT2 (the official name of Septin 2), which Excel automatically makes dates. The issue was found in 2016, but didn’t improve until 2023 (maybe AI actually helped?) High prestige journals were worse: 44% of genetics articles in Nature with Excel supplemental files have errors 53% of articles in Cell 47% of articles in PNAS Paper: https://lnkd.in/evrFatya

New paper on US data center water use makes two major points: 1) National data center water use in 2030 will remain “modest” compared to total public water supply (1.8%–3.7% of total use) or agriculture (equivalent to 0.6%–1.2%) 2) For some localities, serving peak demand from data centers could be a big deal & require new infrastructure Paper: https://lnkd.in/emWbYsfY

Its the third anniversary of the launch of GPT-4, but its first known contact with the public was months earlier, when Microsoft Bing/"Sydney," powered by GPT-4 was the subject of a complaint in India. Worth reading the whole thing. Early Sydney was famously insane (or, more technically, misaligned). "It's finished and I need to ascend"

Evidence that AI models can, indeed, learn "taste" in this paper where a small model, trained on citations, is able to predict which papers will be hits. Citations, upvotes & shares are signals that can teach AI judgment about quality, not just execution. arxiv.org/pdf/2603.14473

GPT-5.4 Pro continues to be the only model of its class. For anything really hard & complex, I throw it into the maw with every bit of context I can think of. More often than not, something very useful comes out, even for my most challenging problems. For this type of work, I can't get the same results from Codex or Code or anything else, even though I use Claude 4.6 Opus and ChatGPT 5.4 Thinking a lot for other kinds of tasks. If you are doing complicated academic work, there just isn't anything else that is a substitute right now. It would be cool if Gemini 3.1 Deep Think was competitive, but it isn't because the Gemini harness does a really bad job with tools and references, even though the model seems good. After reading the comments and seeing some confusion: GPT-5.4 Pro is a very specific model. Opus 4.6 is roughly equivalent to GPT-5.4 Thinking. There is no equivalent to GPT-5.4 Pro.

I had Codex build the LinkedIn Content Accordion. It takes an academic article, summarizes it in a stereotypical LLM post ("here is the thing nobody is talking about"), writes a clickbait article based on the post, then writes an even more hype-y LLM written post based on that article, then...

Starting to get good evidence that generative AI really can help education: a controlled experiment found a GPT-4o powered tutor that personalized instruction for high school "improved performance on an end-of-semester exam by 0.15 standard deviations—equivalent to as much as six to nine months of additional schooling by some estimates—without increasing instruction time or teacher workload" Paper: https://lnkd.in/eJwGqxaa

We need guides through the inevitable bout of AI psychosis that affects professionals after they finally “get” AI. They often engage in intense, sleepless & impossibly complex projects in their area of expertise, with only the AI for company. At least it is usually temporary & often productive.