Igor Babuschkin

We’re launching River AI, a new AI company with the mission to build AI systems that are owned and shaped by you. I’m extremely excited about what we’re going to ship soon. Check it out!

At humans&, we train models from the long-term impacts of their interactions with people. This requires prioritizing long-horizon multi-agent RL. We've developed and are excited to share an open-source, hardware-native 4-bit RL recipe, significantly accelerating training

Sign up for the River API if you want to try it out. We’re open ing up access incrementally.

We are looking for exceptional people to join the team. If our mission resonates with you - please apply.

Congrats to @elonmusk and $SPCX! “Any man who can hitch the length and breadth of the galaxy, rough it, slum it, struggle against terrible odds, win through, and still knows where his towel is is clearly a man to be reckoned with” - Douglas Adams

We are releasing River API, our first product, in early access. The API gives you access to the same battle-tested tools that we’re using internally at River for post-training, reinforcement learning and continual learning. Check it out and let us know what you think!

@kylekosic Thank you Kyle!

Congrats on the launch! Excited to see what comes out of this group

@zeeshanp_ @TheNormanMu Thank you Zeeshan!

@ibab @TheNormanMu Congrats on the launch Igor!

We’re launching River AI, a new AI company with the mission to build AI systems that are owned and shaped by you. I’m extremely excited about what we’re going to ship soon. Check it out!

@eliebakouch Thank you Elie!

@ibab congrats! the mission and vibe of the video looks nice! (also SWE bench and dinner spots is all you need)

@alexgraveley Thank you Alex!

@ibab Congrats Igor!

@jack Thank you Jack!

congrats to @ibab and team!

We raised $250M in Series C funding at a $2.2B valuation, led by a16z. Exa is a search lab organizing the web's data for agents.

@sun_hanchi True

Greg Yang should be regarded as one of the greatest researchers in Deep Learning. 20 years later, the DL textbook will starts with tensor program

@jukan05 There are many factual errors in this post

Why did xAI hand over a 220,000-GPU cluster to Anthropic? The technical backdrop to xAI's decision to hand Colossus 1 over to Anthropic in its entirety is more interesting than it appears. xAI deployed more than 220,000 NVIDIA GPUs at its Colossus 1 data center in Memphis. Of these, roughly 150,000 are estimated to be H100s, 50,000 H200s, and 20,000 GB200s. In other words, three different generations of silicon are mixed together inside a single cluster — a "heterogeneous architecture." For distributed training, however, this configuration is close to a disaster, according to engineers familiar with the setup. In distributed training, 100,000 GPUs must finish a single step simultaneously before the cluster can advance to the next one. Even if the GB200s finish their computation first, the remaining 99,999 chips have to wait for the slower H100s — or for any GPU that has hit a stack-related snag — to catch up. This is known as the straggler effect. The 11% GPU utilization rate (MFU: the share of theoretical FLOPs actually realized) at xAI recently reported by The Information can be read as the numerical fallout of this problem. It stands in stark contrast to the 40%-plus MFU figures achieved by Meta and Google. The problem runs deeper still. As discussed earlier, NVIDIA's NCCL has traditionally been optimized for a ring topology. It works beautifully at the 1,000–10,000 GPU scale, but once you push into the 100,000-unit range, the latency of data traversing the ring once around becomes punishingly long. GPUs need to churn through computations rapidly to keep MFU high, but while they sit waiting endlessly for data to arrive over the network fabric, more than half of the silicon falls into idle. Google sidestepped this bottleneck with its own custom topology (Google's OCS: Apollo/Palomar), but xAI, by my read, has not yet reached that stage. Layer Blackwell's (GB200) "power smoothing" issue on top, and the picture comes into focus. According to Zeeshan Patel, formerly in charge of multimodal pre-training at xAI, Blackwell GPUs draw power so aggressively that the chip itself includes a hardware feature for smoothing power delivery. xAI's existing software stack, however, was optimized for Hopper and does not understand the characteristics of the new hardware; when it imposes irregular loads on the chip, the silicon physically destructs — literally melts. That means the modeling stack must be rewritten from scratch, which in turn means scaling is far harder than most of us imagine. Pulling all of this together points to a single conclusion. xAI judged that training frontier models on Colossus 1 simply was not efficient enough to be worthwhile. It therefore moved its own training workloads wholesale onto Colossus 2, built as a 100% Blackwell homogeneous cluster. Colossus 1, on the other hand — whose mixed architecture is far less crippling for inference, which parallelizes more forgivingly — was leased in its entirety to an Anthropic that desperately needed inference capacity. Many observers point to what looks like a contradiction: Elon Musk poured enormous capital into building Colossus, only to hand the core asset over to a direct competitor in Anthropic. Others read it as xAI capitulating because it is a "middling frontier lab." But these are surface-level reads. Look at the numbers and a different picture emerges. xAI today holds roughly 550,000+ GPUs in total (on an H100-equivalent performance basis), and Colossus 1 (220,000 units) accounts for only about 40% of the total available capacity. Colossus 2 — built entirely on Blackwell — is already operational and continuing to expand. Elon kept the all-Blackwell homogeneous cluster (Colossus 2) for himself and leased out the older, mixed-generation Colossus 1. In other words, he handed the pain of rewriting the stack — the MFU-11% debacle — to Anthropic, while keeping his own focus on training the next generation of models. The real point, then, is this. Elon's objective appears to be positioning ahead of the SpaceXAI IPO at a $1.75 trillion valuation, currently floated for as early as June. The narrative SpaceXAI now needs is that xAI — long the "sore finger" — is not merely a research lab burning cash, but a business with a "neo-cloud" model in the mold of AWS, capable of leasing surplus assets at high yields. From a cost-of-capital perspective, an "AGI cash incinerator" is far less attractive to investors than a "data-center landlord generating cash." As noted above, the most important detail of the Colossus 1 lease is that it is for inference, not training. Unlike training, inference requires far less tightly synchronized inter-GPU communication. Even when the chips are heterogeneous, the workload parcels out cleanly across them in parallel. The straggler effect — the chief weakness of a mixed cluster — is essentially neutralized for inference workloads. Furthermore, with Anthropic occupying all 220,000 GPUs as a single tenant, the network-switch jitter (unanticipated latency) that arises under multi-tenancy disappears. The two sides' technical weaknesses end up complementing each other almost exactly. One insight follows. As a training cluster mixing H100/H200/GB200, Colossus 1 was an asset that could only deliver an MFU of 11%. The moment it was handed over to a single inference customer, however, that asset transformed into a cash-flow asset rented out at roughly $2.60 per GPU-hour (a weighted average of the lease rates across GPU types). For xAI, what was a "cluster from hell" for training has become a "golden goose" minting $5–6 billion in annual revenue when redeployed for inference. Elon's genius, I would argue, lies not in the model but in this asset-rotation structure. The weight of that $6 billion becomes clearer when set against xAI's income statement. Annualizing xAI's 1Q26 net loss yields roughly $6 billion in losses per year. The $5–6 billion in annual revenue generated by leasing Colossus 1 to Anthropic, in other words, almost perfectly hedges xAI's loss figure. This single deal effectively pulls xAI to break-even. Heading into the SpaceXAI IPO, this functions as a core line of financial defense. From a cost-of-capital standpoint, if the image shifts from "research lab burning cash" to "infrastructure tollgate stably printing $6 billion a year," the entire tone of the offering can change. (May 8, 2026, Mirae Asset Securities)

@_arohan_ There is only Training

There is no pre-training, post-training, or test-time training. There are only priors, updates, constraints, and compute budgets. There is only TRAINING. Last several years we shipped the org chart to fundamental optimization science.

sglang is the best inference framework out there. RadixArk was formed to make it even better and to democratize more of the frontier AI stack. Very happy to support the team in their seed round.

DeepSeek V4 by @deepseek_ai just dropped! SGLang is ready on Day 0 with a full stack of optimizations from architectures to low-level kernels. We also deliver a verified RL training pipeline in Miles (by @radixark) for V4 at launch: 1️⃣ Native "ShadowRadix" Design: DeepSeek V4's hybrid attention is complex. Our new ShadowRadix engine is the first to provide native prefix caching for SWA and compressed KV pools, making 1M+ context retrieval seamless and memory-efficient. 2️⃣ High-Performance Kernels: - Flash Compressor: IO-aware fused kernels, 10x faster than naive implementations. - Lightning TopK: High-speed indexing for 1M context in just 15µs. - Integrate FlashInfer trtllm-gen MoE, FlashMLA, and MegaMoE kernels 3️⃣ Rich Features: Speculative decoding, HiSparse, Attention DP/TP/CP and MoE TP/EP, and multi-platform support 4️⃣ Verified RL: The open-source RL pipeline: full parallelism (DP/TP/EP/PP/CP), tilelang kernels, tensor-level checked precision, verified with growing reward. Get started immediately with our out-of-the-box Cookbook 👇 Enjoy! #DeepSeekV4 #SGLang #LLM

@beffjezos We call that scalable oversight

The trick is to have a 150 IQ CEO that hires 160 IQ's that then hire 180 IQ people.

@Rafa_Schwinger I agree with that. Canada is also worth mentioning with the Perimeter institute. The US also has many pockets of diverse research ideas, they just seem small in comparison to the mainstream.

The biggest problem is that the American research ecosystem got too centralized and your proposals had to follow some top 5 big shot or no way to get grants. In comparison Europe has a more heterogeneous system and as a result the breadth of theoretical programs and approaches is higher (eg in quantum gravity and quantum foundations) even if a bit weaker. It is insane for the USA to be such a research monoculture when it is a giant country.

It seems inevitable that LLMs will soon surpass human physicists at creating new theories. The next big breakthrough in theoretical physics will likely come from someone prompting a model.

By the way, if you push today’s LLMs to come up with new knowledge, they struggle noticeably compared to repeating existing knowledge (published papers). So there are still difficulties with strong generalization. This seems like something that will be solved soon though.