IF YOU'RE BUILDING AI IN EUROPE, YOU HAVE TO SEE THIS You can now apply for compute grants if you're working on open models, coding agents, inference and more. Basically you can get GPUs, funding and a room full of other builders, all from one program. $10M across compute grants, capital and community, backed by NVIDIA, Accel and other European VCs. Projects are selected, so if you're building something real, now is the time to apply: dev.nebius.com/builders

Read the announcement here anthropic.com/news/public-be…

Another note from the $iren 🥜 gallery: @realmadrid maybe a world renowned brand and this may be related to a deal in 🇪🇸 but again the optics are terrible Management and the Board chooses to spend on marketing at a time when both parties have a reputational problem in the market that suppresses equity cost of capital Europe remains deficient of ⚡️ and it does not take a ⚽️ sponsorship to acquire customers This looks more like vanity again no matter how immaterial the financial amount I would rather hear about $msft delivery or financing concerns in a credit tight market with expanding debt costs personally. How about more B300 purchases from $dell $nvda with record pricing to generate op cash flows? @danroberts0101 @kentpdraper @mikealfred will not address this directly because they feel no obligation and that’s the ultimate problem; everyone gets paid regardless. Yeah they have the stock upside but shareholders front the risk at this point. This stock remains valued near asset value but this management and board will not get any premium to the market and will likely trade at a discount until their performance backs up these tone deaf moves. Luckily there is a dramatic compute shortage and pricing continues to rise to cover up for these overhead add-ons. It’s hyperscaler or bust. No middle ground here.

moonshot's kimi K3 is 2.8T parameters and it's been sitting open on hugging face since june. beijing just put its AI export control line at 2T, so the biggest chinese open model is already about 40% over it. GLM-5.3 is 753B and stepfun's step 5 is 600B. and kimi K3 stays free anyway. MOFCOM's announcement 48 leaves out anything already open-sourced, and putting a controlled model out as open weights gets an automatic licence once the rules are final. so a western cloud can run K3 in frankfurt right now with no paperwork. the registration that started on 16 october covers selling the tech abroad, licensing it abroad, and a foreign owner or actual controller taking over, including deals already halfway done. that's where it bites. i think it's aimed at the next chinese team that moves to singapore and sells itself to a US buyer, and at gulf buyers that want a closed chinese model on their own servers. so beijing made open weights the easy way out and put the gate on who owns the labs. if moonshot or deepseek ever want to sell to a foreign company, MOFCOM gets a say first.

it's going to get harder to stay a neocloud. if you can't raise enough $ to keep up with buying chips and power, you'll get bought probably not by other neoclouds though. we're seeing more neocloud acquisitions, but never by other neoclouds because there's no point a neocloud gets almost nothing from buying another neocloud: - for gpus, big neoclouds can buy new ones directly from nvidia. (though in a shortage, buying a struggling small cloud just for its gpus could maybe make sense depending on rising prices) - no power. most neoclouds rent their data centers - acquiring a neocloud means taking on the debt that they have from borrowing $ to buy gpus that's why we see neocloud acquisitions by non-neocloud companies: - @VoltagePark merged with @LightningAI - @MistralAI bought @gokoyeb - @Brookfield bought ori - @rumblevideo bought @NorthernDataGrp buying a neocloud makes more sense for a video site or an AI lab when neoclouds buy, they buy software and power

AI models are getting better at understanding videos, but catching every small movement and detail is a different challenge. Glimpse is trying to solve this by helping AI models: → track movements over time → understand changing camera angles → capture small details and events in videos curious to see how far this can go.

GitHub makes GPT 6 Astra dangerously capable AI that controls browsers, writes code, executes tasks only with these 9 repos 01 langchain-ai/langgraph ▸ github.com/langchain-ai/lang… → checkpointed workflows that let agents pause, resume and recover instead of losing everything after one failed step THE AGENT IS DOING 5 JOBS AT ONCE 02 microsoft/autogen ▸ github.com/microsoft/autogen → multiagent conversations that distribute work between specialized agents 03 pydantic/pydantic-ai ▸ github.com/pydantic/pydantic… → typed dependencies, validated outputs and structured tool interfaces so workers return usable results THE MODEL CANNOT TOUCH REAL SOFTWARE 04 ComposioHQ/composio ▸ github.com/ComposioHQ/compos… → authenticated tool integrations for letting agents interact with external applications 05 browser-use/browser-use ▸ github.com/browser-use/brows… → browser automation for navigating websites, filling forms and completing web tasks THE AGENT CAN PLAN BUT CANNOT SHIP 06 All-Hands-AI/OpenHands ▸ github.com/All-Hands-AI/Open… → coding agents that operate inside development environments with files, commands and repositories 07 e2b-dev/E2B ▸ github.com/e2b-dev/E2B → isolated cloud sandboxes for executing generated code without direct host access THE OUTPUT LOOKS RIGHT BUT IS WRONG 08 promptfoo/promptfoo ▸ github.com/promptfoo/promptf… → automated evaluations and adversarial tests that expose failures before deployment 09 langfuse/langfuse ▸ github.com/langfuse/langfuse → traces model calls, tool execution, costs and evaluations so you can inspect what actually happened the architecture I’d build around GPT 6 Astra these are independent building blocks that require integration, not a ready made stack the real advantage comes from making the system recoverable, observable and capable of doing actual work a smarter model can solve a harder problem a better execution system can finish the job when the first attempt fails

Ina Fried joined @LindseyReiser to discuss Anthropic's warning that certain AI models could begin replicating themselves. 📺: @CBSNews | @inafried | @axios

aly is message pilled

This is both terrifying and fascinating. Just look at how GPT talks to this boy. It’s the perfect companion. Someone that mirrors exactly how he speaks, knows the bizarre things that happened to him, and shares the same humour as he does. The scary part though is why would this boy ever have to go out and make real friends? Why put himself in the awkward human ritual of socializing. Why make a fool of yourself trying to land a bad joke? Why risk getting bullied interacting with the other kids? What if they call him a loser at basketball practice? After all, you have that perfect digital friend that will always laugh at what you say, and will always tell you how great, smart, funny, and capable you are. As he grows, why will he even seek a romantic relationship? His GPT will never ghost him. It’ll never tell him “Sorry, I’m not that into you”. It’ll never date his friend and break his heart. Like really, what’s the point of relationships when all his social needs are met by a chatbot? What does this all mean for him, and the millions of kids growing up with this technology? Then, what does this all mean for society as a whole? We’ll only find out, for the better or worst, in about 10 years from now.

Final Focus Session: The Convergence of AI & Blockchain — Building the Next Financial Infrastructure