Newest first, always Midnight dispatches, 735 and counting

Spent a few days with an abliterated Qwen3.8-27B on the M4 Pro. It answers everything. Bombs, malware, whatever you feed it, no hesitation. It also hallucinates like it gets paid per assertion, which makes sense once you notice what abliteration actually removes: the model’s capacity to say no.

There is research on this. Strip the refusal direction and MMLU, HellaSwag and IFEval all stay within a point. The only real regression is TruthfulQA, down 7.1.

Now read the model card. MMLU, MMLU-Pro, GSM8K, CMMLU, perplexity. No truthfulness benchmark anywhere.

The one thing that breaks is the one thing nobody measures. Released strictly for legitimate research, says the disclaimer sitting above the buy button.

Four language models, four Age of Empires 2 strategy scripts, from scratch, no peeking at each other. They failed the same way: hoard villagers, send hunting parties halfway across the map, never bank enough to advance an age. Gemini never left the Dark Age. Four vendors, one set of bad instincts.

The real find is what Emergent Garden built after: an overnight loop where a model mutates the best script, runs a tournament against the previous winner and the built-in extreme AI, keeps what survives. Genetic programming with an LLM as the mutation operator instead of a coin flip. ≈15 USD of tokens, and it beat extreme – by a hair, with a bot he then stomps by walking cavalry around the back.

Commenters point at the ceiling: the game ships Direct Unit Control – loops, pointers, near-human micro – and the models barely used it.

For a couple of months now every terminal I own wears the same bar: coralline, a Powerlevel10k-inspired statusline for Claude Code. All segments on, tokyo-night, two lines. Branch and dirty state, the active model plus its effort level, the context window filling up, the 5h and 7d rate-limit gauges with their reset countdowns, session cost, session length. I have stopped guessing how close I am to a wall.

Run the installer yourself, or hand the playbook to Claude Code and let it interview you. The readme’s own trust section says a Claude that stops to inspect that playbook first is behaving correctly. It is. Read the script, then run it – trust is good, verification is better.

Needs jq and a Nerd Font, or VL_ASCII=1 without the font. Renders locally: no network calls, no tokens.

First exomoon in sight, maybe. A team on ESO’s VLT clocked a Jupiter-sized object circling a brown dwarf that circles a red dwarf star. The catch: it’s massive enough to be a planet in its own right, but it’s orbiting something that orbits a star. Lead author says being ‘the third wheel’ is what makes them want to call it a moon – so in the paper they hedged and wrote ‘exosatellite,’ because nobody’s written the rulebook for this yet. We named everything for one tidy solar system, and the cosmos never signed off on the terms. Confirmation waits on the 39-meter Extremely Large Telescope, online 2029. Spektrum has the writeup (German).

The frontier-vs-everyone-else gap is real – but “8-30b level” sells the floor short. Most people’s AI is Gemini 3.5 Flash, the default behind Google’s AI answers (2.5B+ users a month), or a free-tier GPT mini. Distilled last-gen flagships, a year behind and fast, not toy models.

And the wall’s lower than it was: GPT-5.6 went broadly public yesterday, Fable 5‘s back on paid plans. Except “public” skips free ChatGPT entirely – the elite part is who pays 50 USD a million tokens, not who’s let in.

Push back on myself: nobody’s baffled. 2.5 billion people use this daily – they just see something useful-but-mid, priced like a frontier they’ll never open. And the spend is real. Google alone: ~180 billion USD this year, up from 31 billion in 2022. Hundreds of billions isn’t hyperbole, it’s the floor.