Welcome to #109 of the AI edge.
BTC spent the week stuck between 75-80k, which is the same place it was last week and the week before. The range continues to hold…for now.
Zcash is where the money went. New all-time high this week, up around 160% on the year, now a top ten coin. The "private bitcoin" trade stopped being a meme somewhere around the third digit. NEAR is catching the same bid from the private AI side, which tells you the demand is one theme wearing two hats.
On the AI side, Claude keeps losing ground to OpenAI, especially since Astra shipped. Anthropic's response was not a model. It was a blog post.
Dario's "We Must Pace the Frontier" argues the whole industry should deliberately slow capability gains so safety work can catch up. It's a serious argument. It also arrives from the lab that just got lapped, which is a hell of a time to discover the brakes.
Elsewhere, Meta's Muse hit number one on the App Store. Zuck has now topped the charts with a social network, a messaging app, and an agent. The man knows how to ship to a billion people.
With that, let's get into this week's edition.
The Big Story: The Frontier Model That Couldn't Watch Football
Ask a frontier model what's in a photo and it will tell you, fluently and usually correctly. Ask it to do the same for every frame of a 90-minute football match and the economics collapse.
Score, which runs Subnet 44 on Bittensor, put GPT-6 Astra Ultra through its football evaluation this month and it scored 0%. The rules penalise false detections heavily enough that a model which gets most of a crowded frame right can still end up with nothing. Their estimate for running a full match through Astra or Claude Fable 5.1 anyway is roughly 1,000 times what their own models cost.
Subnet 44 is where those models come from. Independent engineers compete to build small, specialised vision models for narrow tasks, validators score every submission against published criteria, and the winners get compressed down to around 19MB, small enough to run on an ordinary CPU instead of a dedicated GPU per camera. That work already runs commercially through Manako, Score's enterprise arm, across 120+ AVIA sites, 60 Shell and ENI forecourts, Reading FC and Lavance in vehicle washing.
This week they opened the workshop by launching Score Studio, which brings the whole computer vision lifecycle into one place.
The Wedge
Everything in one workspace: Generate and label data, train a model or pull one off the subnet's public track, inspect exactly where it fails, then wire models and logic together on a visual canvas and ship it to the cloud or to a local edge box. Today that work is scattered across notebooks, labelling tools and separate model hubs, and nobody enjoys it.
The part nobody else can offer: If no existing model solves your problem, Model Foundry takes the task, your evaluation criteria and a budget, and lets subnet engineers compete to build it. You approve the benchmark result before you accept the model. Ordinary vision platforms can only hand you tools and wish you luck.
It ships with an MCP endpoint, so an agent can call vision as a tool rather than trying to be one.
The Fine Print
The 0% result comes from Score's own evaluation, scored against rules Score designed. Penalty-weighted tests can be built to make almost anything fail.
Model Foundry is the headline feature and the least proven one. The example Score shows on its own site is a drone crop monitoring model sitting at 50.1% against an 85% target, which is a work in progress rather than a delivered result.
The subnet's public models are only as good as the problems the competition has already covered. Detection, tracking and counting are well served right now.
The models can think about anything. Looking at one thing, all day, cheaply, is the part still up for grabs.
Tessara Watch: The Tool That Wears Out First
Everyone worried about AI board supply has been watching capacity. The tighter number is how fast the tools get used up.
An AI server board is a stack of copper layers with insulation between them and signals have to move across those layers and also vertically through them. That means holes. Once drilled and plated, the holes become vias, the vertical connections between layers. An advanced server board needs tens of thousands of them.
Lasers handle the shallow ones. Deep holes through thick multilayer boards still get drilled mechanically, by a tungsten-carbide bit spinning at around 200,000 RPM. That bit is a consumable. It has a hit count, the number of clean holes it can make before the cutting edge gives out, and at the smallest diameters that can be a few hundred.
Every trend in AI hardware pushes the number down. Boards are thicker, holes are finer, and the low-loss laminates that protect signal quality are harder to cut. So drill demand can grow faster than board volume itself.
It's already showing. At Topoint, high-end coated drills went from 48% of volume in 2025 to 56% in the first half of 2026, with operating margins reaching roughly 28%. Union Tool, which holds more than 30% of the global market, grew first-half operating profit over 80%.
Supply is answering, though. Topoint is going from about 35 million units a month to 45 million by year-end, then 70 million by end-2027. That's why High-Speed PCB still reads Balanced, with drill bits tightening underneath it.
The more interesting question sits upstream. Drill capacity can be built in quarters. Tungsten cannot, and China produces over 80% of it and put carbide powder under export control last year.
Which raises the real question:
If high-end drill demand keeps climbing, what tightens first, precision grinding capacity or tungsten itself?
Lium (SN51) generated nearly $1M in monthly revenue, using the proceeds to buy back and burn $1M of its alpha token while expanding its GPU network to 68 datacenters across 21 countries.
Chutes (SN64) is training an 8B model across 30 distributed GPU hosts with Parallax, achieving 3.9M tokens/sec at roughly $11 per billion tokens.
Prime Intellect's Prime Agent surpassed 20,000 GitHub stars, reflecting growing adoption of its open-source autonomous coding framework.
Oro (SN15) became the first startup in eight years to enter Y Combinator with its token already live, marking a milestone for tokenized startups.
OpenServ AI launched the SERV Reasoning API on Base, offering production-grade AI reasoning with built-in compliance and privacy features.
OpenRoboto (SN80) is now available in the Coinbase app via Aero on Base, bringing its open robotics network to 100M+ Coinbase users beyond the Bittensor ecosystem.
ORO has evaluated 2.5M+ AI agent trajectories in just 101 days, using the data to significantly improve its in-house 4B model while distributing up to $20K/day in onchain rewards to top builders.
🔥 Our Weekly Top Tweets
#1 NEAR Intents Hits New Volume Record
NEAR Intents processed over $300M in volume in a single day—nearly matching the protocol's entire $409M cumulative volume from July 2025, highlighting the rapid growth of intent-based transactions.
#2 Anthropic Says RSI Is Getting Closer
Anthropic revealed that 26% of its AI R&D work is now performed by AI, up from under 1% in February, with around 30,000 agents working internally. The company says these results show how close the industry is to recursive self-improvement (RSI).
Cheers,
Teng Yan & Arvind
And before we close out..
Go deeper with Tessara
This edition gives you the weekly read.
Tessara keeps the underlying research alive: the constraint, the companies exposed, the evidence, and what would change the conclusion.
