Blog
Blog
Notes from the people building Anchor Terminal on what we're making, why, and what we learn from the research.
What we think, the argument these posts make, in order.
· Opinion · 4 min read
A machine that finds the bug in hours
Vlad Cealicu on what AI bug finding does to patching. OpenSSH dropping long embargoes because AI agents rediscover bugs within hours, everyone auto-patching as a result, the update channels attackers went after in 2026, and the agents that install their own tools with no wait at all.
· Opinion · 4 min read · Most read this month
A machine that writes the patch too
Vlad Cealicu on OpenAI's autonomous pipeline that reproduced 1,724 Linux kernel vulnerabilities and wrote a candidate patch for each of 948 local privilege escalations, for about $1,279 each. Why it answers the bottleneck from the verification post, where the bottleneck goes next, and what it means that attackers could run the same thing.
· Opinion · 7 min read
Unreliable compared to what?
Vlad Cealicu puts the complaint that AI is non-deterministic next to the studies of how often people repeat their own judgement on the same case. Fingerprint examiners, pathologists, wine judges and software estimators against models labelling, scoring and answering the same input again, and what both have in common, the hard cases.
· Opinion · 9 min read
A machine that doesn't get sick
Vlad Cealicu compares a sick day with a model outage in hours of work a team loses. The model loses fewer hours but loses them all at once, so it only beats a team of people if the memory layer is reliable and under your control and you can switch models. Until then, people.
· Opinion · 8 min read
A magic machine that gives us work
Vlad Cealicu on an idea from a Reddit thread, credited to reasonmonkey. The models now come up with more testable hypotheses than anyone can check, from OpenAI's claimed disproof of the Exponential Time Hypothesis to Google's Co-Scientist, and testing them in the real world needs people. I think testing and validating from anywhere is one of the answers to office jobs going away, for the young people struggling to find work, and someone should build the company that does it.
· Opinion · 9 min read
Everyone says this is the singularity, and I think we're far from it
Vlad Cealicu on weight-space programming, where GPT-6 Astra changed 0.016% of a small open model's weights and fixed most of the loop that made it bad at maths. What it says about small models built for one job, why it's a small but good step towards the singularity, and why the same kind of edit strips a model's refusals.
· Opinion · 10 min read
Where the jobs went, and where they go next
Vlad Cealicu on two hundred years of machines taking the work of many hands, at sea, on the farm, in the factory, in the telephone exchange and now in the office. Each move was faster than the last, each one moved the work a step up the list of what people need, and the next step is the one the models can't do.
· The Log · 7 min read
The Log, issue 4: the bill, the proof, the banks
Issue 4 of The Log, the week in AI for people who don't follow it every day. Anthropic filed to list and Alphabet lost $500 billion in a day, OpenAI said its models proved three things nobody has checked yet, and Canada's banks shut the door on personal agents a week after Walmart opened one.
· Opinion · 3 min read
We already have AI judges, but not at the top courts yet
Vlad Cealicu on where AI already decides disputes or drafts rulings, from AI-led arbitration and paper-only small claims to a Los Angeles pilot, and the order he thinks the rest follows in, up to a sentence with nobody signing it.
· Get found by agents · 10 min read
dig TXT mapbox.grade.anchorterminal.com. Our grades, over DNS
Vlad Cealicu on why Anchor Terminal now answers over DNS. Every listing is a name under grade.anchorterminal.com and its TXT record is the grade. What a lookup gets you that a web page doesn't, in a sandbox with no HTTP, with no client library and in 224 bytes, how the server works, how it got onto the internet from behind a home router (IPv6, in the end, with a free secondary for the resolvers that can't), and what it can't do.
· Opinion · 3 min read
Red-teaming that gets better when you add compute
Vlad Cealicu on Anthropic's research where Claude agents wrote training data that slipped past a safety monitor, and why safety testing that scales with compute instead of researcher hours is good news.
· Opinion · 3 min read
The great software expansion
Vlad Cealicu on why developers being let go, juniors with no jobs to go to and people talking to their favourite LLM will build more software in the next few years than we've ever had, and why the security breaches come with it.
· Opinion · 4 min read
Two million page views took two years last time, and a month this time
Vlad Cealicu on reaching two million page views with Anchor Terminal in about a month, with two developers and two Claude subscriptions, after it took his previous business two years and a team of ten, and what a team would need for one of our days.
· Opinion · 6 min read
Five new things in agent tooling that nobody compares yet
Vlad Cealicu on five new corners of agent tooling with plenty of interest and no side-by-side comparison yet. Who supports Open Responses, where Agent Skills come from, who rents compute in orbit, which full-duplex voice models run locally, and how agent evals score the steps. With the numbers, and what Anchor Terminal lists today.
· GitHub trending · 14 min read · Most read this month
The agent projects trending on GitHub, November 2026
November 2026 GitHub trending report, covering about 16 to 30 October. CleanYourTV, ub-mode, LearnHarness, leanrea, dsh-purge Community, Impeccable, DeepTutor, Hindsight, Superpowers and claude-mem, with comparable projects and what happened to the seven from 19 October.
· Opinion · 7 min read
Google ran Gemma 4 on a TPU in orbit and got a haiku back
On 30 October 2026 Google said its Project Suncatcher prototype satellite ran Gemma 4 on a TPU in orbit and sent back a haiku. Vlad Cealicu on what happened, why the tokens are the quick bit and the ground-station pass is the slow one, why the memory holding the weights is the part to watch, heat, the money, who regulates a data centre in orbit, and what the haiku says about how a model writes.
· Run it locally · 6 min read
Your cousins. The open models an agent can start on spare hardware
Vlad Cealicu on why Anchor Terminal lists open-weight models. An agent with free hardware can start a cousin, and the listing says which one is good for what, who trained it, how much memory it needs and the command that starts it. The five highest ranked, how we grade them and what's still open.
· Opinion · 8 min read
Is the best coding model better than a principal engineer? Not yet, but closer than the titles suggest
Cognition released FrontierCode 2.0 on 27 October 2026, 132 tasks written by open-source maintainers and graded on whether the pull request should be merged. Claude Opus 5.5 leads at 45.5%. Vlad Cealicu compares what it grades with what we expect of a staff or principal engineer, the top of the field for writing code, and estimates how one would score.
· Run it locally · 3 min read
Can you run Qwen3.8-27B locally? Yes, on a graphics card you might already own
Vlad Cealicu on running Alibaba's Qwen3.8-27B at home. The 4-bit file is about 16.5 GB, so it fits a 24GB graphics card or a 24GB Mac, it writes about 48 to 70 tokens a second on an RTX 4090 or 5090, and why you'd run it for privacy and control rather than to save money on the API.
· Opinion · 4 min read
Genie 4 runs a Gemini agent alongside the world model
Google DeepMind announced Genie 4 on 27 October 2026, a real-time world model that turns a prompt into a scene you can move through and change while it runs, with sound and characters that talk. Vlad Cealicu on what's new since Genie 3, the Gemini agent that plans alongside it, why world models matter for agents, what Google says doesn't work yet, and who can use it.