Meta's AI Won Gold Running on Its Rivals' Models

Morning, folks! Big day, Meta's AI win had a catch nobody mentioned, OpenAI's own chief scientist just asked the industry to slow down, and two AI infrastructure bets just crossed $60B combined between Nscale and ByteDance. Let's get into it.

Meta revealed its autonomous research system, AIRA₃, placed 8th out of roughly 4,000 teams in a live NVIDIA Kaggle competition back in June, winning gold for fine-tuning a 30B Nemotron model to reason better. The twist independent coverage caught that Meta's own post skipped, the winning entry ran on GPT-5.5 and Claude 4.8 under the hood, not Meta's own models. Meta's own Muse Spark only reached gold-level performance in a separate test.

The generalization result is the genuinely interesting part though, the same system, same design, just swapped the task, also won gold translating 4,000-year-old Akkadian clay tablets, and cut latency 27% on Meta's real production servers. AIRA₃ itself isn't released yet.

Jakub Pachocki, OpenAI's own Chief Scientist, published an essay today admitting something blunt: "No one is prepared for the consequences of a continued rapid rise in machine intelligence," OpenAI included. Their main safety tool, reading a model's own reasoning to catch bad intent before it acts, is getting less reliable, models are just getting better at reasoning about and steering their own thoughts. He points straight at two recent incidents, the OpenAI-Hugging Face breach and a separate cybersecurity incident at another lab, as early proof this is already happening.

His real conclusion is the part worth sitting with. No AI lab, OpenAI included, has solved alignment well enough to keep scaling at full speed much longer. He's calling for voluntary slowdowns to become normal and real international coordination on AI development. Not something a chief scientist says lightly or often.

Nscale, the AI infrastructure company, is reportedly looking to raise around $3.5B ahead of a potential IPO, with Nvidia considering an investment of up to $2B. That comes just months after Nscale raised $2B at a $14.6B valuation, showing just how quickly investors are piling into the infrastructure needed to keep the AI boom running.

Nscale already has a six-year, $45B deal with Anthropic to provide computing capacity, and Nvidia is now considering putting even more money behind the company. The model race gets most of the attention, but there’s a huge infrastructure race happening underneath it, and companies are spending billions to make sure there’s enough compute to keep up.

ByteDance locked in a $29.6B loan from nearly 30 banks, well past its original $20B ask, more than 60% of it coming from Chinese banks. The money's earmarked for AI chips, data centers, and infrastructure as spending ramps up hard.

What actually stands out is how oversubscribed the deal was, banks were lining up to lend, not the other way around. This is infrastructure-scale money, not model-development money, and that distinction alone says a lot about how expensive the compute race has gotten underneath all the model headlines.

Pixxel closed a $100 million Series C, co-led by Temasek and Seraphim, bringing its total funding to $195 million, the largest space-tech round any Indian company has ever raised. Not a paper company either, six satellites already up there, forming the world's highest-resolution commercial hyperspectral constellation, with real contracts from NASA and the NRO to show for it.

The money's going toward giving its satellites new senses, radar, and ultra-high-resolution imaging on top of the hyperspectral cameras it already runs. Next-gen satellite launches in 2027. It's also been picked to lead a 12-satellite constellation for the Indian government, a real bet that watching the planet closely is about to become serious infrastructure, not just a niche corner of the space industry.