Signal #63: Albania Owns 0.08% of the Frontier

A Balkan government owns a piece of a frontier AI lab, and this week that lab shipped its first model. The terms are stranger than the headline, and they say something uncomfortable about what buying into the frontier gets you.

Then three moves from the labs you use every day. The open-weights crown changed hands twice in eight days and DeepSeek was not in the room. OpenAI published the attacker model it built to break its own agents, which you should read if you took my advice over the last month. And Anthropic gave Claude away to America's teachers, which is not charity.

Ready to transform artificial intelligence from a buzzword into your personal revenue generator?

HubSpot’s groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age.

Inside you'll discover:

Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.

Get Your Guide

On July 15 Thinking Machines Lab released Inkling, its first model: open weights, 975B total parameters with 41B active, a million-token context, trained on GB300s. The lab says plainly that Inkling is "not the strongest overall model available today, open or closed." The Albanian government owns a piece of the company that built it. Tirana put in $10M, roughly a billion lek, joining the a16z-led seed at a $12B valuation, and it had to amend the 2025 state budget to write the cheque.

Now the terms. Ten million dollars at a twelve billion dollar valuation is about 0.08% of the company. The Information reported that founding shareholders hold stock carrying 100 times the voting power of ordinary shares, and that Murati's board vote outweighs every other director combined. Albania bought economics, not a say. The $50B round the lab chased last November collapsed in January without a deal, so the $12B mark has not moved in a year.

Read it as a bet, not a trophy. Every other government in the region answered the AI question with a strategy document; one of them bought equity in the frontier and got a shipped model. The position is small, illiquid, non-voting, and held at a price the private market just declined to raise. It is also the only claim on frontier upside anyone in the Balkans holds. Watch the next mark: if Thinking Machines raises above $12B, Tirana will have out-returned most of the region's venture funds on a single cheque.

The open-weights frontier moved twice in eight days. Inkling landed on July 15. A day later Moonshot AI released Kimi K3, 2.8 trillion parameters, the largest open-weights model ever shipped, benchmarking against Opus 4.8 at $0.30 per million input tokens, with weights landing on the 27th. The company that invented this category was absent from both.

DeepSeek's newest model is V4 Preview, dated April 24, and its own documentation index lists nothing after it. That is eighty-four days, and the only July entry is a notice that the legacy model names retire on the 24th. In January 2025 DeepSeek repriced the entire industry over a weekend. Today Kimi K3 is 1.75 times its size and Moonshot owns the headline. Being cheap and open stopped being a moat the moment everyone else could afford to be cheap and open too.

The lesson travels well beyond DeepSeek. If your product's edge is a price point, someone with more compute takes it from you on a Thursday. Check what your own moat rests on this week, and if the honest answer is "we are cheaper," you have a quarter, not a business. Then download K3's weights when they land and price your own inference against them before a competitor does it for you.

On July 15 OpenAI published GPT-Red, an attacker model it trained to break its own systems and says it will never deploy. On a held-out prompt-injection arena it beats human red-teamers, 84% attack success against their 13%. It hijacked a live vending-machine agent, repricing stock to fifty cents and cancelling a customer's order, and it pulled data out of a Codex CLI agent.

I have spent three issues telling you to hand work to agents, so here is the other side of that ledger. An agent with tool access is a system that reads text written by strangers and then executes it. OpenAI's own numbers: its "fake chain-of-thought" attack worked more than 95% of the time against GPT-5.1, and under 10% against GPT-5.6. The newest model fails on 0.05% of direct injections, six times fewer than OpenAI's best model four months ago. That curve is real progress, and it also means most of what you deployed before this year was defenceless.

Version numbers matter more than brand names now. Go find out which model sits behind each agent you have wired into a live system, because "we use ChatGPT" tells you nothing about whether you are running the one that broke 95% of the time or the one that breaks once in two thousand. Then cut the permissions on anything touching money or customer records down to the smallest set that still does the job.

Anthropic launched Claude for Teachers on July 14: free premium Claude for verified US K-12 educators, open for sign-up through June 2027. It arrives with Claude Code and Cowork included, connectors into Canva Education, MagicSchool, Diffit and Brisk, curriculum standards for all fifty states, an FERPA-aligned data agreement, and an open-source repository of teaching skills.

Free tiers for teachers are an old move, but the shape is new. Anthropic is giving away the model, publishing the domain expertise, and wiring itself into the standards bodies and the tools teachers already open, which puts the switching cost in the workflow rather than the subscription. OpenAI is buying enterprise seats. Anthropic is buying the habit of a generation that does not have a budget yet. Both are distribution strategies, and only one of them is free to the user.

The playbook transfers to any regulated vertical with a professional body and a compliance requirement: give the product to the practitioners, open-source the expertise, clear the regulator's paperwork before anyone asks, then let procurement catch up. If you sell into healthcare, law, accounting, or education anywhere in this region, that is the shape of the competitor arriving in two years. Worth doing this week: the skills repo is public, so read how Anthropic encodes a vertical and copy the method for yours.

Four tools to install or test this week. Marketing sat this one out, nothing shipped worth your time.

Productivity: ChatGPT will finally let you find your own work. OpenAI added unified search to the sidebar on July 14, one bar spanning past chats, projects, images and documents, filterable by type, on every plan. If you have been treating ChatGPT as disposable, your account just became an archive you can query. Go find the half-finished spec you abandoned in May and reopen it.

Design: Canva Code turns a prompt into an editable site. Canva shipped Code 2.0 on July 15 to every tier, free included: prompt a website, app, or interactive piece into existence, then edit the result in the normal Canva editor. The point is not the generation, it is that a non-technical teammate can change the copy afterwards without touching code. Rebuild one landing page from a prompt and hand it to someone who has never opened a repo.

Sales: Clay moves into the chat window. Clay launched Clay for Reps on July 16, exposing its data and GTM workflows through MCP inside ChatGPT, Claude, and Microsoft 365 Copilot, with per-rep permissions and credit budgets so RevOps keeps control of the spend. Your reps stop switching tools to enrich an account. Wire it into Claude and research a live account from chat before your next pipeline review.

Dev: Claude Code got a browser. Anthropic's week 28 release gave Claude Code on desktop a sandboxed built-in browser that opens docs, designs, and any site, and clicks through pages the way it already handles your local dev server preview. You decide whether sessions persist. Point it at a third-party API doc, have it read the auth flow, and let it write the integration with no copy-paste in between.

Where the region and the wider market gather this autumn. The good hotels go first, so book the ones worth the flight now.

IJCAI-ECAI 2026, Aug 15-21, Bremen. The joint international AI research conference, heavy on machine learning, language, and vision. Go if you want the frontier research a season before it reaches your stack.

Startup Fair, Sept 17, Vilnius. The Baltics' largest startup event, with a pitch battle putting 40+ startups in front of 200+ investors.

Bits & Pretzels, Sept 28-30, Munich. Europe's founders' festival, roughly 7,500 attendees including 1,500+ VCs, capped with an Oktoberfest finale.

How to Web, Oct 6-8, Bucharest. Eastern Europe's biggest startup conference, drawing leaders from OpenAI, Meta, and Google plus 250+ investors. The Investors Summit runs on the 6th. The most CEE-relevant stop on the calendar.

World Summit AI, Oct 7-8, Amsterdam. The 10th edition, Europe's premier AI-specific conference, anchoring World AI Week.

Wolves Summit, Oct 14-15, Kraków. Tight CEE corporate dealflow, roughly 100 investors and 250 founders, plus the European CVC awards.

Web Summit, Nov 9-12, Lisbon. The continent's largest tech gathering, 70,000+ attendees from 160+ countries.

Slush, Nov 18-19, Helsinki. The most investor-dense flagship in Europe, the one to attend if you are raising at Series A or deploying it.

Next edition soon,

Çelik