Welcome to #106 of the AI edge.
BTC is still stuck in the mud, chopping between 62k and 66k for what feels like the tenth week running. Nothing to trade, nothing to see.
The action was all in equities, and it got ugly. Semis and the AI names spent the past few weeks bleeding out, and the marquee casualty was a name you know.
Leopold Aschenbrenner, the 25-year-old ex-OpenAI researcher who spun his 2024 "Situational Awareness" manifesto into a $45B AI focused fund, watched the whole thing unravel in a matter of days.
He was running his AI infrastructure longs at roughly 4x leverage, the trade turned against him, the margin calls landed, and by the end of the week his entire public book got force-sold to Ken Griffin's Citadel at a discount, all while his wedding was going on. Forty-five billion down to around ten. The AI prophet got liquidated on the AI trade.
And then the market did him dirty. Two sessions after the fire sale, the S&P printed a fresh all-time high, its first record since June, led in part by the same chip names that had just been left for dead.
If only Leopold could have held on a little longer. This has to be the villain origin story they make movies about. Already can't wait for the film.
Ask Tessara
Leopold’s week is the whole point.
AI infra longs at 4x, force-sold into the washout, then the same chip complex ripped two sessions later. The question was never “AI good or bad.” It was which physical constraint was binding, who was exposed, and whether the market had already priced it.
That question is why I built Tessara.
These past 9 months I’ve been heads down on our research terminal for the AI buildout. It runs on our proprietary living model of the physical constraints in AI. We start with the constraint, not the ticker: what’s binding, who benefits, who gets squeezed, and whether it’s in the price.
We just opened the research workspace. Search the evidence. Challenge the view. Ask Tessara a hard question on a name or a bottleneck and get an answer with receipts, plus what would break the thesis.
Is HBM still tightening, and what would prove that wrong?
What binds NVIDIA next?
Is the memory squeeze priced in?
I would love for you to go poke at it.
With that let’s get into this week’s edition.
The Big Story: Matching Trained Models On A Fraction Of The Data
Most models show up to a benchmark having already studied for the exact test. That studying is pretraining, the long, expensive phase where a model chews through mountains of data before it ever faces a real task.
Rei Labs (a project we covered last year in our research), just put out something that walks in cold and still competes.
Adapt-1 Preview is a research build of a system they call ADAPT, a memory-and-reasoning core that learns during a task instead of before it. It is not a transformer, it does not generate tokens, and it does not keep a language model in its decision loop. You hand it a Domain, a plain declaration of what it can observe and what a good outcome looks like, and it builds up task state as the data streams in.
The Wedge
On a partially observed card game from POPGym, a standard reinforcement-learning benchmark suite, Adapt-1 learned across 1,550 live steps, averaged around 0.51 reward, and edged the strongest baselines over its final episodes. Those baselines each trained for 15 million steps, so it reached a comparable score with roughly ten thousand times less interaction.
On BOP-Ask, a robotics reasoning benchmark, the same core started cold and improved across three passes. By the third it led four of seven comparable metrics against models fine-tuned on a 33.8-million-example corpus, carrying no equivalent training of its own.
Every decision keeps its evidence, its version, and its reasoning, so results can be traced rather than taken on faith. Rei also shipped a public guide for reproducing the numbers, which is the step that lets outsiders check the work instead of clapping for a chart.
The Fine Print
The standout number, the POPGym card-game result, comes from a single run, not enough repeats to rule out luck, and it is stacked against baselines that were scored a different way. Better to treat it as a promising signal, not a settled result.
On RoboSpatial, the spatial-reasoning test where a model answers questions about where objects sit in a scene, Adapt-1's own learning was switched off completely, and most of the real work fell to an outside object detector and some fixed geometry math. That result says more about those bolted-on tools than about the core itself, and it still came in about seven points behind the current leader.
Bigger picture, this is genuinely interesting but very early. There is no formal paper yet, no live demo to poke at, and no outside evals to confirm the numbers hold. Fair enough for a first preview, and since this is only the initial launch, there should be more to look at soon.
This thing does not top every column. The part worth watching is that a system holding no pretraining can keep pace with ones that ingested millions of examples, a quiet case that raw scale is not the only path to competence.
Prime Intellect launched Prime Agent, a self-improving coding agent harness for long-running autonomous tasks that achieved 95.5% on ARC-AGI-3, surpassing the human-expert baseline.
Bittensor activated Root Reborn, replacing passive root staking with validator-managed subnet baskets that compete on returns and direct buy flow toward selected subnets.
MoonPay introduced PayBox, a non-custodial wallet that lets Claude and ChatGPT users trade assets and complete real-world purchases through natural language using x402.
Minos (SN107) co-authored an OpenAI field report on agentic scientific computing, highlighting the HelixForge engine behind its decentralized mutation detection network.
Arkham integrated Coinbase's x402 standard, allowing AI agents to pay for real-time onchain intelligence with USDC at request time.
IOTA (SN9) launched Orion-16B, a live distributed training run spanning three continents using up to 256 consumer GPUs.
Leadpoet (SN71) entered a pilot with Dropbox and a channel sales partner, bringing its AI sales prospecting platform to enterprise teams.
Virtuals Protocol launched Hyperboost, extending token momentum after graduation with 14 days of trading and creator rewards funded from reserved token supply.
🔥 Our Weekly Top Tweets
#1 Base Is Becoming the Agent Chain
79% of all registered onchain AI agents now live on Base, while 47.3% of all agent transactions on the network flow through Virtuals Protocol, showing where the agent economy is increasingly taking shape.
#2 Frontier AI Runs on Gaming GPUs
EngyAI got the 2.8T-parameter Kimi K3 running on just 80 consumer RTX 5090s, serving the largest open-weight model without scarce HBM memory or specialized networking. It's a step toward making frontier AI accessible to startups, labs, and universities.
Cheers,
Teng Yan & Arvind
Quick recap: I also publish Tessara Research, focused entirely on the physical AI buildout and the companies exposed to its constraints. 1 - 2 research memos each week.
