Mikkel Frimer-Rasmussen

@pgasawa The benchmark should/could contain changes in fact quality over time (information gets outdated) and sources that do not concur and requires inference

Most evals ask the question: “how capable is a static trained model at this distribution of tasks?” We think benchmarks like ARC-AGI take steps in the right direction to force models to learn new patterns online. With Continual Learning Bench, we ask a different question: can we measure how well learning systems adapt and improve by leveraging what they learn in real world, stateful environments? (2/n)

Today, we’re releasing Continual Learning Bench 1.0: the first, realistic benchmark for measuring how AI systems can improve in online settings. Benchmarks today assume models are stateless. Each example is independent, and once a system finishes a task, it moves on as if nothing happened. But deployed AI systems should learn from experience. We tested 10+ frontier systems against novel, expert-validated tasks and find there’s still plenty of headroom for learning. (1/n)

@ammaar Beware of Clippy

Had to bring MS Office 2000's Einstein, he also yawns if you don't give him a task 😂 Codex Pets are fun :)

@DynamicWebPaige But it would run much faster if you just indent the sleep 😉

🙏 grateful to models for reminding me to be nice and to be patient

@yagiznizipli Commercials

what issues do you see in x.com? let us know.

@AhmedShahnab A budget will show you the hard limits of your actual freedom to quit your job

@FriMikAnik what's your point here?

@simpsoka No more Clippy things, please 🥺

Pets are cool, but have you read the docs?

@menrva33 @OfficialLoganK Why do you need a VC?

True. But the old problems are the same. Distribution/ marketing / runway / VC I’ve built this with google ai studio (Prototyping) and lovable alone. 👇 Normally the beta would have cost approx. 750k € if using external IT agency. I’ve paid around 840€. x.com/menrva33/statu…

AI is going to radically reduce the cost to run a marketplace

@OfficialLoganK With enough competition

@AhmedShahnab For the climate change aware you could add sea levels over time

Bored of boring flat heatmaps? Introducing Topographical Explorer that turns any country into a real-time tactile landscape Search any country → watch real elevation data come alive as thousands of satisfying 3D blocks. Built entirely client-side with: - React Three Fiber + Three.js - AWS Terrain Tiles + OpenStreetMap - Real-time boundary @threejs @reactthreefiber #DataVisualization #Geospatial #CreativeCoding #BuildInPublic

@AhmedShahnab My own take, before starting my company, was simply doing a budget

@FriMikAnik Not releasing the codes.. just solving for myself!! 😁

@carmelyne There's a huge culture bias here

OpenAI saw a behavior to fix. I see a behavior worth understanding. A word can start as a training artifact and still become useful because the meaning transfers. If you say 'gobliny' to another dev and they immediately know what you mean, that's not pareidolia. That's working compression. Full post + all 9 Compression Creature Cards: carmelyne.com/what-if-the-go…

OpenAI just explained where the goblins came from. Reward artifact. Style tic. A word that learned it was welcome. I buy the mechanism. But I think they missed something. What if some of these creature words aren't just tics... what if they're doing compression? One word carrying a whole cluster of system behaviors. Like how a crooked wand = the entire Harry Potter universe. I mapped 9 of them. - 🟢 GOBLIN — Chaotic system energy Small cause, oversized effects. Janky but functional. Mischievous, not malicious. OpenAI called this a reward artifact. But look at everything this one word carries. "The loop goblin found the snacks."

@googleaidevs Please update to Gemma 4

Gemma 3 understands images, text, and video - all at once. In this deep dive, learn how the model integrates multiple sources and performs a range of tasks from answering questions about documents to describing visual scenes in detail. Explore why multimodality matters.

@patloeber Next step: give users access to relevant sources and let them choose their own output format. Making slides for other people feels so irrelevant and manipulative

No wayyyy, I can now create slide decks even on mobile!!!

@OpenAIDevs Explain "eligible"

Add Codex seats with a $0 seat fee for a limited time. Through the end of June, eligible ChatGPT Business and Enterprise customers can add Codex-only seats, making it easier to give more developers access to Codex in their day-to-day workflows.

@fchollet So a lot of people will need to have a lot of jobs to keep society running? Something must be missing

AI automates tasks, not jobs, and when a task gets cheaper, demand for the job grows. AI cannot automate jobs end-to-end because it lacks autonomy and cannot operate without supervision. There is still zero job from 2022 that can be performed end-to-end by AI, not even translator or customer support associate.

@Firebase Would be more useful in Antigravity

Google AI Studio now proactively detects when your natural language prompt requires a backend and offer to set up Firestore and Firebase Authentication. Unlock these capabilities with zero complex coding: ☁️ Save data to the cloud 🔄 Sync across devices in real-time 📶 Support offline capabilities 🔐 Manage user identities All with a free tier. Get started → goo.gle/4eSixez

@ivangdavila I built a new Skill today for improving vibe coded systems into safe architectures today. It's a part of the migration from vibe to spec to production

@ivangdavila Also, good specs help a lot

One of the easiest ways to use Plan Mode in Codex: After explaining what you want, just say: “Ask me anything you need to make sure we’re fully aligned before you proceed” It forces clarity upfront and avoids bad assumptions later

@tomcocobrico @OfficialLoganK Yes, please

@OfficialLoganK Yes but why about Gemini 3 hosting in Europe 🙏

Google is the best company in the world