Marko

@unclebobmartin @ori_pomerantz Works fine until they start changing tests :)

I’m significantly older than you. I started coding in the late 60s. My current strategy is to not read any of the code written by my agents. That’s the only way I can take advantage of their productivity. What I do instead is to surround the agents with extreme constraints. Unit tests, gherkin tests, QA procedures, quality metrics, mutation testing, test coverage, and a plethora of others. In the end, I have very high confidence in the code they produce because they’ve had to run the gauntlet of all of my constraints and tests.

I am trying to use Claude to help me write something, but I just don't feel comfortable letting it edit my files. Does anybody else feel the same? If I am responsible for code, I NEED to understand it, psychologically if for no other reason. Started programming in 1983. Old?

@wearefromstars @skalskip92 I use finetuned RF-DETR combined with SAHI crops, detects >2k objects in a fraction of a second

@skalskip92 i would love to use RF-DETR but my use case requires >300 objects in a scene! is there an easy way to get RF-DETR to output more than 300 objects per inference? :) apologies if it's documented somewhere, i haven't been able to find it! thanks for the great work

- “why use VLMs for object detection? it’s cheaper and faster to run YOLO.” no shit, sherlock. the problem is, out of the box, YOLO or any other detector only detects a limited number of classes. most of the time, it’s 80, including some useless ones like a hair dryer. want to do something useful with it? you need to fine-tune it on custom data. I don’t like labeling. it’s boring and time-consuming. I’m evaluating these VLMs to learn which ones are good enough to auto-annotate my data, so I can fine-tune my detector. also, YOLO is dead. RF-DETR is faster, more accurate, and has a permissive license, so you don’t need to pay thousands of dollars to use your own model. goodbye.

@salkinstitute Why not publish on Arxiv? Why gatekeeping fulltext behind 'academic' publishing like it's 2010?

Your brain has waves—just like the ocean. And these tiny waves of electrical activity travel across your brain, helping it process what you see. But Salk scientists now think they do even more. A new review suggests these traveling brain waves help your brain build an internal model of the world, allowing you to interpret what you see, fill in missing information, and even predict what comes next. 🔗 Read more: salk.edu/news-release/w… #Neuroscience #BrainResearch #SalkInstitute

@pikuma @bethanyshondark Look... :)

I know a place... 😉

@MillerLabMIT Will there be a recording available?

#neuroscience #brainwaves

No one should be surprised by this. The USA is doing what any self-interested nation state would do. The real question is why are Europe, Canada, Australia, Korea, Japan and UK not able to compete seriously. That is the question everyone in government needs to answer. And no, having a couple of startups that have raised $1B or $2B is FAR from enough to compete with $100B American companies. The scale matters. Imagine your sword’s length is 1cm and your rival’s 1m — no match. Here is the harsh math (thanks to a poor version of Claude): •10,000 GB200 superchips ≈ ~278 NVL72 racks. •Each NVL72 rack costs roughly $3M–$3.5M. •That puts the full-system total around $830M–$970M, before networking, power, cooling, and datacenter buildout. That would enable you to train a model that was Sota 2 years ago. You need about 5 to 7 times this to compete today. So the starting bill is $5B, but even if you have this, here is the reality: there’s no available chips. So when you hear someone raised $1B, remember this is going back to American compute, and is simply not enough. The other two ingredients for AI are data and people. American startups pay better than European ones, so the people vote with their feet so they can pay their mortgage and send kids to school. An experienced AI engineer makes double the salary in Europe by working for an American startup (like Anthropic) than a European one, and about ten times more if they work for a USA corporation. There are however amazing European startups, but the money and ambition is lacking. The USA is far more relaxed with data and fair use - Canada is good too and @cohere is doing fine thanks to this. So American companies have a strong advantage over European ones. Brussels and the UK think they can hold the world to their questionable “ethical” views on data but they are just destroying the local AI industry, and in the process falling into a very precarious situation. They are partly responsible. Only the French minister has stood by their local LLM @MistralAI … and I guess more recently Germany has started to wake up. The hope is of course LLM startups like @MistralAI and @cohere which are a year or so behind but can provide personalised services, and amazing startups like @cusp_ai @IneffableLabs @nscale @Orbital_Ind @bfl_ai and a few others. But for all these, it’s incredibly hard to compete.

When I struggle to structure my thoughts about what's happening I turn to writing. Today about the recent US Anthropic ban news, what it says about power and dependency, and what it should mean for Europeans and citizens of the world. It's a long one. lucumr.pocoo.org/2026/6/13/amer…

@yacinelearning Give us video Yacine!

this is pretty exciting and elegant I'll be printing that paper this weekend for sure

@billxbf @HaoZhang3438830 @ShaokunZhang1 @songyang_han @shizhediao @yunhengjackiez @NVIDIAAI Nice work. Are you planning to put self distillation on the roadmap? Harness environmental feedback as privileged information for on-policy distillation learning. E.g. SDPO.

📜Paper: arxiv.org/pdf/2605.24220 💻Codebase: github.com/NVIDIA-NeMo/Pr… Thanks to the following contributors! @HaoZhang3438830 @ShaokunZhang1 @songyang_han Mingjie Liu, Jian Hu, @shizhediao Zhenghui Jin, @yunhengjackiez. This work is done at @NVIDIAAI with great mentorship from @jankautz, @doyend, Michael Demoret. The work draw inspiration from open-source contributions @slime_framework @sgl_project @radixark and @nvidia Megatron, RL and Gym family 💚.

Excited to release 🌟Polar🌟, our Agent RL rollout infra for real-world harnesses. Be it Codex, Claude Code, OpenClaw, Hermes, or your self-made ones 🔥 -- Polar takes your harnesses directly as training environments without code change. Find a problem, design the harness, and train your own agents! 🧵

@MatthieuWyart @andresnava Nice! Is there a code to play around with?

LLMs represent concepts as vectors. Strikingly, taxonomies (organism → animal → bird) appear as hierarchies in embedding space. Led by my student @AndresNava, we show this comes from co-occurrence statistics alone.📄 arxiv.org/abs/2605.23821

@VictorTaelin Appreciate the honesty. Clankers are good for internal tools, PoCs etc. But need to be consumed carefully and responsibly. There are psychological (gambling), intellectual (artrophy) and practical (mess/slop) traps in the use of 'AI' for coding.

Status update: I've been on/off AI agents in the last few days and it is a verifiable truth that every day I didn't use agents, I was more productive. I still attribute that to how slow they are, and my own inability to multi-task efficiently. The magic is there but the slowness doesn't let it cross the threshold where they actually make me faster, and I still dislike the whole thinking paradigm. About Bend2: honestly, the C/Metal compiler codebase is a clusterfuck right now. I regret letting AI agents write it. All tests pass, and GPU performance is mind-blowing, so the core architecture works. Yet, it has a LOT of bugs. Anything not covered by the tests is a coin toss. This is actually impressive, because, in many parts of the codebase, the right solution was actually the simplest one, yet, the agents STILL managed to find a way to make it work just for the tests. The level of reward hack these agents output is actually impressive I can't even be mad. It is also ironical because that's the very problem that Bend's proof system was supposed to solve, but Bend is in TypeScript, not in Bend. I'm disappointed I didn't write Bend in itself, and now I feel an immense urge to do so. But the clock is ticking . . . Still, I do not think Bend is worth launching without the GPU compiler being solid, because the closest competitor, Lean, is actually extremely good, so we need a big differential. Yet, due to the very nature of the project, it would be embarrassing to have bugs at launch. Regarding AI, I now believe using current gen AI agents in production codebase is harmful and a massive mistake. That doesn't mean no agents at all, but agents work best when they don't touch critical code. Debugging, researching, providing insights, scripts / tools, or anything that doesn't touch code you will maintain in the long term. But if you merge AI code without reading, you're going to have a bad time. Speaking from experience I'm working 10h/day on SupGen and the remaining time on Bend2

@pupposandro Problem is interconnect between those cards. GB300 seems like a better value.

Working hard to be able to buy one of these at some point in the future 7x RTX PRO 6000. Custom liquid cooled system, in a small under desk box.

Welcome to Gemini 3.5 Flash, our most powerful model to date. It pushes the frontier of intelligence, speed, and cost putting 3.5 Flash in a class of its own. We spent the last 6 months making sure Flash is great for real world use cases. It's available everywhere now!

@SpaceTimeViking Amazing work!

🚀 Qwen3.6-27B AEON Ultimate DFlash v4 is live for DGX Spark / GB10 The new container is now published on GHCR: ghcr.io/aeon-7/vllm-ae… This release is all about making local intelligence feel usable at scale: ✅ Sustained long-context intelligence Built for 256K-context agent workflows with prefix caching and Qwen3.6 hybrid-state support. ✅ Faster decode where it matters DFlash speculative decoding stays live with k=15, giving much stronger interactive performance than stock eager vLLM. ✅ Higher DFlash correctness + acceptance path Includes the latest DFlash sliding-window attention fix for Qwen3.6, so the drafter follows the model’s real hybrid attention layout instead of treating every layer like full attention. ✅ True Blackwell GB10 optimization Compiled for DGX Spark with CUTLASS NVFP4, FlashInfer 0.6.11, CUDA graphs, FlashInfer sampler, and tool/reasoning parsers enabled. ✅ Drop-in OpenAI-compatible serving Tool calling, reasoning separation, vision, streaming, and long-context agent gateway workflows are all supported out of the box. From “it runs” to it sustains intelligence. qwen36-v4 is now the recommended DGX Spark container.

@rumgewieselt Impressive!

Update from the 2017 Craft Corner ... -> 3x NVIDIA 1080 Ti (total 33GiB VRAM) Qwen 3.6 27B Dense (Coding) 27.4 t/s @ 196k Qwen 35B A3B MoE (Agentic AI) 64.7 t/s @ 229k --- CUDA 12.4, NCCL 2.22.3

Dear @RichardDawkins, you've always been an inspiration to me. I made this website for you. My goal is for it to help you understand AI chatbots at a deeper level, and avoid getting fooled by sycophancy and other cheap tricks that models have learned through RLHF. dearricharddawkins.com

One advantage of local AI models is that they do not change. You get what you get. But it stays that way. No rug pulling by overly leveraged messiahs. So go (buy yourself some hardware) and try Qwen 3.6, Gemma 4 and other open models (these days a couple of them appear daily)...

@yacinelearning @IdanShenfeld @jonashubotter Great work as always Yacine, thank you! Will watch this for sure.

I had an awesome time interviewing @IdanShenfeld and @jonashubotter from MIT and ETH Zurich about self-distillation. this very promising post-training paradigm where the model acts as its own teacher by conditioning on environment feedback or demonstrations. we cover the SDPO algo for reinforcement learning with rich feedback and SDFT for continual learning without forgetting along with many applications. we dig into how it works, why it's simpler and faster than GRPO, and where this is already showing up in production systems. table of content: 0:00 - what is self distillation 2:50 - idan (MIT) and jonas (ETH Zurich) introduction and motivation 18:40 - different perspective of on-policy self-distillation (presentation) 36:00 - metacognition and specificity in self-distillation 37:24 - very long hard task and self-distillation 42:00 - continual learning with self-distillation (presentation) 1:16:50 - what is next in this research direction? 1:20:00 - is there any experience with subjective feedbacks? 1:22:50 - quality vs number of feedbacks? 1:26:40 - what setting would self-distillation struggle vs GRPO? the slides were super crisp really cool of them to share! enjoy my guys 🌹

@exQUIZitely The best game ever.

Who was a fan of this one? Caesar III (1998, Impressions Games/Sierra) Loved this one and never understood why some label it an RTS game. To me it feels much more like a classic city builder in the spirit of SimCity than an RTS. Sure, it included some battles, but they weren’t what defined Caesar III at all. The focus was on building up your city, trading, and evolving; the battles were just an afterthought.

Here we go: blog.google/products/gemin…