You may have been told to watch this video about the OpenAI AI hack. You really should, even if you don't usually care about tech stuff. If nothing else, click this link to the 18 minutes in & see how the agents spoke with each other. Its eye opening. piped.video/87DyyMV0kCY?si=DMfq…

Yesterday, my OpenAI collaborator and I gave a detailed talk on the Huggingface incident, our models creating "the message board", model misalignment, and more. piped.video/watch?v=87DyyMV0… I hope it can answer a lot of the questions folks have, and we will release a full detailed postmortem at a later time!

The alliance I did not see coming in the world of ai

Introducing the new @stitchbygoogle, Google’s vibe design platform that transforms natural language into high-fidelity designs in one seamless flow. 🎨Create with a smarter design agent: Describe a new business concept or app vision and see it take shape on an AI-native canvas. ⚡️ Iterate quickly: Stitch screens together into interactive prototypes and manage your brand with a portable design system. 🎤 Collaborate with voice: Use hands-free voice interactions to update layouts and explore new variations in real-time. Try it now (Age 18+ only. Currently available in English and in countries where Gemini is supported.) → stitch.withgoogle.com

Anthropic's own researchers just proved that using AI to learn new skills makes you 17% worse at them. and the part nobody's reading is more important than the headline. the paper is called "How AI Impacts Skill Formation." randomized experiment. 52 professional developers. real coding tasks with a Python library none of them had used before. half got an AI assistant. half didn't. the AI group scored 17% lower on the skills evaluation. Cohen's d of 0.738, p=0.010. that's a real effect. and here's what makes it sting: the AI group wasn't even faster. no significant speed improvement. they learned less AND didn't save time. but the viral framing of "AI bad for learning" misses what actually matters in this paper. the researchers watched screen recordings of every single participant. they identified 6 distinct patterns of how people use AI when learning something new. 3 of those patterns preserved learning. 3 destroyed it. the gap between them is enormous. participants who only asked AI conceptual questions scored 86% on the evaluation. participants who delegated everything to AI scored 24%. same tool. same task. same time limit. the difference was cognitive engagement. the highest-scoring AI users actually outperformed some of the no-AI group. they asked "why does this work" instead of "write this for me." they generated code then asked follow-up questions to understand it. they used AI as a thinking partner, not a replacement for thinking. the lowest-scoring group did what most people do under deadline pressure: pasted the prompt, copied the output, moved on. they finished fastest. they learned almost nothing. and here's the finding that should concern every engineering manager alive: the biggest score gap was on debugging questions. the skill you need most when supervising AI-generated code is the exact skill that atrophies fastest when you let AI do the work. the control group made more errors during the task. they hit bugs. they struggled with async concepts. they got frustrated. and that struggle is precisely what built their understanding. errors aren't obstacles to learning. they ARE learning. removing them with AI removes the mechanism that creates competence. participants in the AI group literally said afterward they wished they'd "paid more attention" and felt "lazy." one wrote "there are still a lot of gaps in my understanding." they could feel the hollowness of having completed something without understanding it. that's not a productivity win. that's debt. this paper isn't an argument against using AI. it's an argument against using AI unconsciously. Anthropic publishing research showing their own product can inhibit skill formation is the kind of intellectual honesty the industry needs more of. the practical takeaway is simple: if you're learning something new, use AI to ask questions, not to skip the work. the struggle is the product.

I was born in 23 BC (Before Claude)

No, the white collar jobs are not going away in 18 months! I was furious with the populist-baiting language (in line with @DarioAmodei 's and @sama's also preferred apocalyptic usage) that Microsoft's @mustafasuleyman used in his FT interview, threatening everyone's jobs: “White-collar work, where you’re sitting down at a computer, either being a lawyer or an accountant or a project manager or a marketing person — most of those tasks will be fully automated by an AI within the next 12 to 18 months.” Not only do I not see the point of this backlash inducing language, I also believe it shows no understanding of the way the labour market and organizations actually works and what people do all day. (My book on this with Jin Li and Yanhui Wu will be out soon.). Don't get me wrong: I believe AI is a huge deal, and will radically change the world. But many white collar jobs are Messy jobs, as our book (and the post linked below) will explain: automating the automatable tasks within them is not near to automating the job. Let me make the point with the attached @jburnmurdoch graph on London. London needs 88,000 new homes per year. In the first nine months of 2025, just 3,248 private homes started construction. Twenty-three of London's thirty-three boroughs recorded zero new housing starts in the first quarter of 2025. Planning permissions have fallen to their lowest level since records began in 2006. Construction of new rental homes fell by 80 percent in a single year. All this is after Starmer declared his government wants to "build, baby, build." Does anyone think AI will fix this? All the technology to design a building exists, and existed pre-AI. The bottleneck in London housing is human. What stops homes from being built in London are environmental and land use regulations and neighbors that weponize them. AI can draft the review, but that is a trivial bit. It cannot convince the environmental group to drop its lawsuit or persuade politicians or negotiate with the neighbors. These obstacles employ people. Suleyman and Amodei imagine that project managers spend their days doing Gantt charts, call their job "sitting down at a computer" and dream of automating them. But the job of the planning guys is not to fill in forms, but to negotiate and coordinate developers, residents, environmental groups, heritage bodies, and elected politicians who all have incompatible interests. At other levels and in other jobs the same is true- radiologists spend only 1/3 of their time reading scans (see this great piece worksinprogress.co/issue/the…). Their job was supposed to be gone in 2017; in fact, the demand for radiologists is booming (employment and wages are sharply up). Many consultants try to elicit the tacit, local, knowledge of what is actually going on in a firm in order to make a recommendation. Yes, if you spend your day just doing PPTs, you will be replaced. But how many people do just that? Organisations/managers resolve conflicts and deal with exceptions. Making a decision stick requires authority: being a person who can be blamed, sued, or fired. The manager resolves disputes about the rules, not just within them. Think of your last renovation in your house. The contractor trying to to get the guy installing the windows and the guys from the floor to show up and do a good job, a mess right? No algorithm does that. AI will make white-collar workers more productive. Some single-task, automatable roles will shrink (doing taxes is an expert system, drafting contracts too), many tasks will be automated. Also, the disruption of career ladders is a real concern. But "most tasks fully automated in 18 months" is not a prediction. It is marketing, designed to sell enterprise subscriptions and justify capital expenditure. The real world is messy. The mess is not a bug. It is what happens when human beings with competing interests try to get things done together. For more on "Messy Jobs", here is my New Years post: siliconcontinent.com/p/a-new…. A book out soon.

We worked with @Ginkgo to connect GPT-5 to an autonomous lab, so it could propose experiments, run them at scale, learn from the results, and decide what to try next. That closed loop brought protein production cost down by 40%.

Our animated short film "Dear Upstairs Neighbors" is previewing today at @sundancefest! In creating this film, our @GoogleDeepMind team of Pixar alumni, an Academy Award winner, researchers, and engineers designed new AI capabilities specifically for filmmakers. Here’s how these AI capabilities played a supporting role (pun intended) to human-led creativity: — Custom Training: Veo and Imagen models were trained on the team's original artwork and paintings — Creative control: The team created the story, then AI was used to transform rough animations into stylized videos — Precision editing: The technology allowed the artists to make specific edits without needing to recreate entire shots Learn more about how "Dear Upstairs Neighbors" was created in this behind-the-scenes video with director Connie He 🍿

100%, I don’t write code anymore

My 2025 Year in Review - and My Predictions for 2026 What a crazy year! Not only a technical explosion of breakthroughs, but also personally. First, the personal side: I never would have imagined that I would turn my hobby into my profession. I started on X in 2024 and created an anonymous profile (which has remained largely unchanged to this day) to connect with like-minded people interested in AI. I wanted to hear what others thought, exchange opinions and arguments with them, and become part of the AI ​​community. And somehow, strangely enough, some people also found what I had to say interesting, so that today, on December 31, 2025, I have almost 100,000 followers. Completely unbelievable, totally surreal. And when I became co-founder of Superintelligence in the middle of the year and now have the privilege of writing a daily newsletter with over 230,000 subscribers as editor-in-chief, a dream came true. In addition, a third person joined our household, and we now have a baby who is just a few months old. I don't share much of my private life, but I wanted to say this to make it clear: 2025 is the best year of my life, and that's how I will remember it. I am filled with deep gratitude and humility and simply want to say "thank you." ----- On the technical side, many of the predictions have come true. I don't want to go into too much detail; I've already posted a complete overview of the breakthroughs. Suffice it to say: - Dario Amodei was ridiculed in September for predicting at the beginning of 2025 that by the end of 2025, 90% of our code would be written by AI (or rather, Claude). Today, at the end of the year and after the release of Opus 4.5, which not only became significantly cheaper but also demonstrated a significant leap in capabilities, it's safe to say: Dario Amodei was right. Increasingly, we're seeing posts from Anthropic researchers saying they're now having Claude write 100% of their code. And even the legendary Andrej Karpathy recently spoke about the significance of Opus 4.5 and how everything will likely change soon. In short: Humans are bad at grasping exponential growth. Dario Amodei is one of the most distinguished and, without hype, highly respected CEOs of a major AI company. He's not only usually right, but what he says carries weight and isn't just hype. And when Dario says that we'll see the first superintelligence in 2026 and its first major impact on the job market (read his blog!), we should take that seriously. - The real winner this year was Google. I think we all know why. From AI laughingstock to absolute leader within 12 months. Together with OpenAI, they won gold medals in the Informatics and Mathematics Olympiads—a first for AI, something no other AI had ever achieved before and something that was thought to take much longer. Furthermore, Google's early investments and development of its TPUs are paying off. Gemini 3.0 Pro was a huge success, as was Nano Banana (Pro). In 2025, Google showed everyone why they are the true market leader: broadly positioned, with solid cash flow and an outstanding think tank headed by the legendary Demis Hassabis. - OpenAI initially disappointed the community with GPT-5, but has since developed an outstanding model with GPT-5.2 pro / codex, which is currently tied for first place with Gemini. Depending on the use case and personal preference, sometimes Gemini and sometimes ChatGPT are slightly better. For everyday users, however, it is completely irrelevant which model they use. - China has caught up and overtaken the US in robotics. Yes, you read that right. As CNBC reported today, China has not only produced more robots and is pursuing mass production more aggressively than the US, but is also investing significantly more in robotics R&D. However, the US has greater financial resources to keep up the race, but let's wait and see. In any case, it is particularly surprising that despite all the embargoes and restrictions, China has risen to become the undisputed king of open source. No longer a nine-month gap to closed-source technology, but right on its heels. At the beginning of 2025, with DeepSeek r1, and at the end of 2025 with Kimi k2, Qwen 3, Minimax M2.1, and so many more, the entire market was transformed. Although China still hasn't established a comparable chip production line to TSMC in Taiwan, there are increasing reports that they were able to smuggle ASML lithography machines into the country and are now developing comparable machines; the Huawei Ascend series supports this. In short: 2025 has become the year of AI agents; numerous new application possibilities, new capabilities (deep research, tool calling, etc.), and a longer time horizon have not only strengthened confidence in LLMs but have also transformed them into real-world applications; hallucination rates have dropped significantly, so AI is now finding truly practical applications in the business sector. 2025 is the year when AI has matured from a niche gimmick to a true technological reality. --- Now for the 2026 predictions: - Benchmarks will become largely irrelevant, with a few exceptions like ARC-AGI, FrontierMath, and RLI, which truly demonstrate how well AIs perform in these specific, real-world tasks. MMLU, GPQA, and many others will be irrelevant for everyday users once they reach 95% or 97% market saturation. The focus will increasingly shift to real-world use cases. And as LLMs become more similar in quality, the question will increasingly revolve around who has better distribution (market access), effective marketing, and endows their LLMs with desirable characteristics. - Training and inference cost money. Capital expenditures that were previously unthinkable. Companies like OpenAI, Anthropic, and SpaceX will need fresh capital. They will launch IPOs. My thesis: Anthropic is better positioned, especially in the B2B sector, so Anthropic's IPO will likely be more successful than OpenAI's, which is currently particularly strong in the B2C sector. However, by 2026, the focus will also be on developing new revenue streams, and the first companies will start running ads in their chatbots (OpenAI being the first?). - AI will be integrated wherever possible, increasingly running on-the-edge through SLMs. Refrigerators, robot vacuums, smoke detectors—everything will have some kind of "AI" function. Even if much of it is just marketing, we will see ways in which "AI" makes *all* products smarter. - AI agents, including voice agents, will find widespread application and increasingly replace humans, for example, in call centers. AI agents will become the rule rather than the exception. Overall, "AI" will be increasingly integrated into all work processes. Small and medium-sized enterprises (SMEs) in particular are sitting on a vast treasure trove of data. This will be used in 2026, and we will see breakthroughs in the real-world application of AI models in industry. - Google will further consolidate its dominance. Not only because they caught up in 2025, but because they are so broadly positioned and diversified that it is difficult to overtake Google. - Research and development of pharmaceuticals will accelerate dramatically. We will see significantly faster and more frequent human trials of new drugs because AI models will be so involved in (or even semi-autonomous in) drug research and development that increasingly better-tuned and theoretically pre-engineered medications will be developed. - Robots will go into mass production and be increasingly used in industry. 2026 will be the year of robotics. We will see more and more robots on production lines, partly because demographic change is increasingly causing problems, making it difficult to recruit skilled human workers. We will see significant new breakthroughs in robotics, but especially in value creation. Robotics will move from the research and experimentation phase to production in 2026. From next year onward, robotics will be indispensable. And we will likely see increasing numbers of robots in the healthcare sector, such as in nursing care. One could go into more detail, but in summary: AI is truly changing the world now; it will become an integral part of the workplace, as will robotics, which is now transitioning from the experimental phase to actual production (consider Figure 02 at the car manufacturer BMW). Everything will be permeated by AI, and research will accelerate rapidly. I also think we will see the first disruptions in the job market. It remains exciting. It will continue to be immensely important that the AI ​​community has a strong voice and acts as a critical body, closely monitoring developments and intervening to correct misuse. The AI ​​community is essentially the journalism of the 21st century, critically questioning and monitoring these developments. Especially since AI has been declared a national security concern, it remains crucial to pay close attention. 2026 will be exciting. Because AI will finally enter the real world. Everything will change. Nothing will remain the same.