π¨π³πΊπΈ The model doesn't decide who wins. The agent harness does.
JIT-Agent auto-generates task-adaptive harnesses for any LLM. DeepSeek-V4-Flash beats GPT-5.6 on DeepSearchQA (+9.1) and OdysseyBench (+4.3).
GLM-5.2 gains up to +20.2 points.
80 hours of brain data from one person just set a record for noninvasive AI decoding.
LibriBrain100 drops 100 hours of MEG, 8x more than the next comparable dataset.
It hits state-of-the-art on word classification, the stepping stone to brain-to-text.
The BCIβ¦
SoftBank is buying humanoid robotics at a 40% discount.
In talks for a majority stake in OpenAI-backed 1X at a $6B valuation β down from the $10B 1X wanted in 2025.
1X sought $1B. Raised less than half.
Masa is betting the physical AI layer is now cheap enough to own.
π¨π³ DeepSeek is about to be worth $74 billion.
The lab that cracked the AI race on a shoestring is closing a ~$7B raise at a CNY 500B valuation β pre-IPO.
That puts Liang Wenfeng's outfit in the same zip code as OpenAI's early rounds. Nobody priced this in. π₯
The 'AI bubble' thesis just collided with a $96.2 billion quarter. Nvidia's revenue is up 106% year-over-year, its Q3 forecast is $108 billion, and every other major headline this week pushes the same direction: the AI infrastructure cycle is still accelerating, not cooling.
7. The next ceiling isn't silicon. It's volts. If you want to know whether the cycle breaks above $108 billion next quarter, watch utility interconnection queues, not benchmark leaderboards.
Sam Altman just told Time he'll have 'AGI' by the end of 2026 β roughly four months from now. The data-center chief responsible for the compute is reportedly already gone. That gap is the whole AI story right now.
6. The power balance: the 'AGI by 2026' claim was supposed to keep OpenAI ahead in the narrative cycle. Instead, it may have accelerated the talent drain to labs that move slower and ship steadier. That's the real race now β who keeps the operators, not who claims the date.