meet @mostik_ai! what happens when you put 12 PhDs in one room for four months? first place on the ARC-AGI leaderboard, which I can't say much about while the competition is still running. and this, which I can. everyone's arguing about whether open models will catch up to frontier models. we think it's the wrong question. here's the one we pose: why does a frontier model have to generate your answer at all, when the only thing you need from it is the reasoning? we do this by enabling models to communicate in latent space. through our protocol, hidden states pass straight from a frontier model into a small one running on your infrastructure -- no text between them, and neither model is fine-tuned. two models from different families, sharing reasoning, both left untouched. how do we know it works? we tested it on a setup where a 753B model reads the problem, and a 4B edge-class model writes the answer. with this approach, we get results 80% as accurate as the frontier model, but at 20x faster performance. we're committed to preventing frontier model lock-in and are already partnering with inference providers to accelerate open-weight adoption. we've done this between 15 of us, in four months, 12 PhDs and a Fields medalist, backed by @generalcatalyst WIRED has the first external account of the company and the work: wired.com/story/russian-star… full writeup, the setup, and all the numbers: mostik.ai/read-more

Amin Vahdat live at Transition-AI 2026 with @CatalystPod. Google's Chief Technologist for AI Infrastructure - the man in charge of Google's $175–185B 2026 CapEx and the one who's said compute capacity needs to 2x every 6 months! Key takeaways: Training → Inference cascade: -Frontier GW training clusters have a ~1–2 year useful life; then capacity cycles to serving -Inference doesn't need GW scale — <100 MW can do useful work on large models -"Entering the age of inference" is now real as agents explode DC footprint -Speed-of-light latency starts to matter as models get faster — geography becomes UX + reliability -"a medium number of medium-sized data centers, augmented with a small number of large ones" (few large training clusters, inference can be mix of 10s to 100s of MW) Reliability reframe -4-nines isn't intrinsic: "we should be thinking about lower reliability power delivery overall" -Do you make the this trade: 4-nines at half capacity, or 2-nines (3.65 days downtime/yr) at 2x capacity? Customers "very often" pick 2x Behind the meter -Google actually prefers grid-connected capacity = BTM is a bridge, not a destination -BTM is about a different latency: time-to-delivery of capacity -Bridge sources: turbines, gas, mobile generation. Permanent: solar, wind, nuclear -Stranded BTM? "I'd love to have that problem." Energy is the limiter Co-design = capability -Google co-designs across Gemini ↔ software ↔ TPU ↔ rack ↔ DC ↔ power ↔ building -A few percent at each interface compounds into real advantage Bottlenecks -Won't force-rank chips/power/labor/EPC: "10am it's labor, noon it's power, 2pm it's chips — every day" -YoY efficiency is real: this year's capacity would've cost ~1.2x last year Building fungibility is dead, purpose-build is back -25-yr buildings vs. 5–6 chip generations. Disk rack vs. GPU rack = ~100x watts/sq ft and widening -Old world: build for fungibility (compute wasn't dominant cost). New world: "this is a GPU building, that's a TPU building" -Density wasn't maxed before because flexibility was worth more. That's flipped One of the most important pods of the year from an AI leader at one of the most important companies at the center of it all! Great job @shaylekann

TPUs are underrated