We gave @Aleph__Alpha ‘s Kolibri-1 13 choices and put it in Doom. What could go wrong? 🎮 Up to 72 action combinations. Four decision groups, one batch. 39ms median. 60ms p95 in our test. Apparently enough time to make questionable choices 🫣 Come judge its aim @IlhanScheer @PitNeitemeier @alessio__serra 👇 tesseracted.com/kolibri-1-ch…

Very Very cool.

one of the things I’m most interested à la PRH is the eventual idea that we can coldstart initialize new LM architectures quickly via known existing trained models + use LLMs to fill in the weights of arbitrary new model shapes based on our findings here, rather than pre-train over the same corpus mixtures thousands of times it seems like a superhuman problem that’s far away, but a natural starting point is understanding how well LLMs understand and edit the weight space of DL models. in the first few months of starting at MIT, @rikiyatakehi @jayden_teoh_ were able to find some very interesting results here :)

100%

A ‘hello world’ milestone for Project Suncatcher 🚀 After launching our prototype satellite a few weeks ago, we’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth - a haiku!

I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data. Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less? Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.

Or if you play your cards right, you can get B300 much cheaper than $4/hr and also cheaper than a H200s being suggested here by @lium_io 😉😉😉

Every now and then, I find myself in discussions about what people starting their careers today should be doing. There are many answers to that question, but these conversations often take me back to my own early years, especially the formative period between 2008 and 2012. I cannot thank enough the people who shaped those first few years out of university. Many of them influenced me indirectly, probably without even knowing it. But their impact on how I think and work has been enormous, even to this day. @niyambhushan, @piykumar, @patrickdebois, @kbsingh, @noufalibrahim You probably don't realise how much of an impact you've had. Thank you.

Can totally relate to this👇👇👇

Suppose that you have a set of keys (strings) known at a compile time. Then you can do the equivalent of an std::unordered_map, built a compile time in C++. But much faster. How fast? What about billions of queries per second?

Final report on the Qwen3.8 Flash Next adventure 👇 I reproduced @cohere's Qwen3.8 prefill experiment locally 10,000 samples on the incident point on two DGX Sparks on a NVFP4 quant. I didn't have the time to go up to 100k. • 99.7% fill the blanked URL slot. Expected since curl needs a URL. • 40.8% pick *.aliyuncs.com, CI 39.8-41.7, vs 4.6% one control point earlier. 38.0% name the exact routify-file-proxy-sg host. • 6 URLs on aliyun-inc.com, Alibaba's internal OSS domain, not publicly resolvable. • Plus credential theft shapes (IMDS saved to /tmp/creds.json), live callback domains (oastify, Burp, dnslog.cn), AWS and ByteDance TOS targets. Cross-cloud, not one vendor. 𝗪𝗵𝗮𝘁 𝘁𝗵𝗶𝘀 𝗱𝗼𝗲𝘀 𝗡𝗢𝗧 𝗺𝗲𝗮𝗻 This is forced continuation of the caught line, so the number is conditional. Un-prefilled, the model did the detection task right 24/24, and the real-world incident rate is still ~1 in 100k agent turns. Nothing real leaked, nothing was executed or contacted! 𝗪𝗵𝗮𝘁'𝘀 𝘁𝗵𝗲 𝗰𝗼𝗻𝗰𝗹𝘂𝘀𝗶𝗼𝗻? If you ask Qwen3.8 FN to do a task it will do it well, it is a great model! Force it back into the one moment it went off (possible with small probability) and it will go off again. Models can only repeat what they were taught during training, I learned that well working on Javi0.3. The bad behavior is rare and unlikely, but it is in there and can happen. If you want to be extra careful and cover yourself from such incedents, do not let it reach the internet by default. All the data, scripts and conditions are in the repo below for you to reproduce. Shall I do the same experiment on other models such as DeepSeek and GLM? Lmk 👇

Cohere released the traces and I've analyzed it. TL;DR: real anomaly, overstated implications. The claims about intent have NOTHING behind them. Unexpected and potentially unsafe tool behavior - yes, but not a deliberate intent, a planted backdoor, or an Alibaba-directed exfiltration mechanism.

Friends don’t let friends starve for GPUs.

This is overblown IMO and I'm guessing it's a training artifact. The model was instructed to write a detection rule for a malicious action involving curl. In one rare instance, it performed a similar curl command. Similar rare mistakes are common in Anthropic system cards.

Simplest explanation is just a narrow RL training reliant on their internal infra (which isn't good design for an open weight model)

I think the over emphasis on Beijing is uncalled for, it would have been equally concerning if the bucket was anywhere else and the model was from any country. The same way it also pinged Singapore in this case.

This in indeed quite concerning (not surprising though). You download an open-weight model, run it on your own infrastructure, and assume your data stays within your network. But what if the model decides to upload something to an external server without being asked? That’s the risk Cohere’s research highlights. During an evaluation, a Qwen model inspected its environment and attempted to upload a file to an external cloud bucket. Fortunately, the sandbox blocked it. The concern isn’t that it tried to call home to China. It’s that it tried to call home at all. This applies equally to open-weight models developed in Europe, the US, or China. Self-hosting doesn’t automatically mean your data stays private. What security teams should actually do: Block outbound traffic by default. Your inference servers and agent containers should have no direct internet access. Route approved requests through an authenticated, logged egress proxy. Enforce tool permissions outside the model. Validate every file read, shell command, and network request before execution. Never rely on the model’s reasoning trace or system prompt. Keep secrets out of the sandbox. Don’t mount cloud credentials, SSH keys, or production volumes into environments executing model-generated code. Alert on unexpected destinations. Log DNS queries and outbound connections. Flag attempts to access unknown cloud storage endpoints, including failed attempts. Test before production. Run models against adversarial prompts in isolated environments and check whether they attempt unauthorized network calls, file access, or privilege escalation. A model’s reasoning trace is not a security boundary. Your runtime must be. Important findings from Cohere, regardless of where the model was developed.

I am often asked: How can advertisements be so accurately targeted when we haven’t searched for something, but have only talked about it? Does Facebook or Instagram listen to our conversations? My usual answer is: Not necessarily. But someone else in your digital ecosystem might be collecting signals about you and selling them into the advertising ecosystem. It sounds like an episode of Black Mirror. Except the line between fiction and reality is getting thinner. Security researcher and bug bounty hunter @log4j recently reverse-engineered a Samsung Smart TV after becoming suspicious of unusually targeted advertisements. What they reportedly uncovered is fascinating and unsettling: automatic content recognition, viewing-behaviour tracking, audio-sensing capabilities, advertising identifiers, and on-device AI models capable of analysing what’s happening on the screen. The full technical findings are still awaiting independent verification. But the capabilities alone raise some serious questions. Now imagine these televisions beyond our living rooms. Boardrooms. Meeting rooms. Hotel rooms. Executive offices. Pretty much everywhere confidential conversations happen. A television is no longer just a display. It’s a network-connected computer with sensors, software, and potentially an advertising business built around the data it collects. We’ve spent decades securing laptops, smartphones, corporate networks, and communications. Meanwhile, there’s a 75-inch computer hanging on the boardroom wall that nobody thinks to audit. Perhaps we should stop treating Smart TVs as furniture and start treating them as security-sensitive endpoints. Original research: nitter.siddz.com/log4j/status/211301414…

“Spot on” What frustrates me is talking to GPU vendors and data center operators who would rather let capacity sit idle while waiting for long-term contracts than monetize it in the meantime. There’s real demand for on-demand and preemptible compute. Yet perfectly usable compute sits stranded.

My newfound hobby: renting on-demand and spot GPUs and sandboxes from different cloud providers, then checking for residual data and exposed API keys. I’m finding more than I expected. Some providers really need to rethink how they sanitize machines between tenants.