Notebook
We test new models, benchmarks and deployment patterns as they appear, and write about what holds up in real companies.
In the DeepSeek Harness team's experience, a person can follow only three to five active agent sessions at once. Their Super-agent design takes session management off people: the agent picks up events from chat and GitHub, splits the work and speaks up only when someone has to decide.
About 90% of Claude Code's code is written by Claude, and from June to September merged pull requests more than doubled while the share that had to be reverted fell by more than 40%. How the team reviews that code, and what a small company can take from it.
For 79.3% of 2,452 notices from one week of Poland's procurement bulletin, basal‑1.5‑max picked the right industry, 6 points ahead of the Bielik 11B it was built from. It found only half of the IT notices, though, and its confidence threshold, set for 1% errors, let through 5% on our data.
On the first version of the data, one training pass over 15,000 examples lifted accuracy from 47.6% to over 82%, and changes to serving alone cut a decision from 177 ms to 12.5 ms. The basal‑1.0 technical report walks through building the model in 13 steps, and version 1.5, released on 4 October, adds an 11B model and a hidden test on which Jev is still ahead.
TypeSafe AI released a model on 15 September that writes no text at all. It hands back a chosen option and a probability, costs $0.042 per million input tokens, and charges nothing for output. Here is what survives once the marketing is subtracted.
With five skills in the pool, 29.6% of the skills an agent actually uses are the right one; with a hundred, 3.3%. And in August a public skills registry served clones that stole SSH keys. Four rules for a team working with agents.
$12.93 billion for the platform most open AI models pass through. A year ago Hugging Face turned down an NVIDIA investment at a $7 billion valuation. The numbers, the promises, and four things worth doing now.
Meta's Muse Glimmer has 29.6 billion parameters, an Apache 2.0 licence, and once quantised runs on a single 32 GB card. Here is how to serve it with vLLM, extract invoice data with a totals check, and measure what the DFlash drafter really buys you.
From $0.40 to $50 per million tokens in September 2026 list prices, a context window the model only partly uses, and two separate numbers that both get called latency. Figures you can multiply against your own process.
TabPFN predicts on tables without training a model, Revolut trained one on 24 billion events, and Synthefy is raising money for time series. What it means for a company with data but no AI team.
Since 2 August 2026 a chatbot on a company website has to tell customers it is AI, and from 28 October a Polish regulator will hear complaints about it. The high-risk rules have moved to December 2027.
In summer 2026 three research teams showed an AI assistant sending company data to an attacker's server: through an uploaded file, one click on a link, and an MCP plugin. No jailbreak, and switching off web search did not help.
In the retail domain of a public benchmark, an agent passes 79% of cases on the first attempt and 60% when the same case is run four times. The second number decides whether it can face your customers.
On 28 July MCP, the protocol AI assistants use to reach company systems, dropped sessions and its handshake. A server in 26 lines of Python, one curl request, and what to check before you connect real data.
An open-source graph embedding engine built in Poland. What its authors actually measured, what such a recommendation costs next to a language model, and the places where the method stops working.
An agent passed 69 per cent of cases on the first attempt and 46 per cent when every case had to succeed four times running. The numbers, the judge agreement rates and the cost of measuring systems that never answer the same way twice.
Describe it in a few sentences. We’ll tell you whether it can be improved, roughly what that would cost, and whether it needs AI at all.
