Independent AI research on agent reliability, evals, memory, and systems. Read more ↓
Joined December 2025
- Tweets4.2K
- Following57
- Followers633
- Likes11.2K
OptiQ Code scored 82.9 on the Harness Leaderboard's Terminal-Bench mini + DeepSWE mini tasks, state of the art there. The best harnesses on the board, pi and mini-swe-agent, get 82.2 on the same 29 tasks with GLM-5.3-Flash. All nine are in the chart.
mlx-optiq.com/code?utm_sourc…
The same idea exists in the open for npm: Frame's `deps --vulns --backport` takes the advisory's upstream fix commit, applies it to the installed version as a patch-package patch and shows the diff first, no upgrade needed. Apache-2.0.
github.com/lambdasec/frame/t…
AI turns new vulnerabilities into exploits within hours. Upgrading to patch is too slow.
In real codebases, @SocketSecurity can patch 25 to 50% of critical CVEs today without upgrading a single package. Every patch is human-reviewed. Critical patches are free.
socket.dev/features/patches
New OptiQ quant: Groot-19B-A3B-OptiQ-4bit, GrootTeam's mixture-of-experts model, 13.5 GB on disk, takes images. Capability Score 78.37: GSM8K 94.4, BFCL-V3 92.5, HumanEval 89.6, IFEval 77.1, MMLU 70.6, HashHop 46.0. Free: pip install mlx-optiq
mlx-optiq.com/models/groot-1…
A PTS steering vector for Qwen3.5-0.8B cuts stock literary words ("hum", "whispered", "testament"): answers using one fell from 57% to 17% on 50 held-out prompts. Prompting "Write plainly" only got to 44%. No fine-tuning. Plainer prose, not better prose.
huggingface.co/LocalLLaMA/Qw…
The Ramayana in 30 seconds, told with characters from free browser games 🪔 Happy Diwali.
mythology.games/?utm_source=…
Frame 1.1.3 turns a finding it proved reachable into a runnable test that drives the app's own code to the sink, in Python, JS/TS, Java and C#. Then --check-fixed re-runs that test on the fixed code: it passes only if the fix removed the flow.
github.com/lambdasec/frame/r…
Mandiant: average time from disclosure to exploitation fell from 63 days (2018-19) to 5 in 2023. Verizon's 2024 DBIR: ~55 days to fix half of critical known-exploited flaws.
Frame checks which vulnerable deps your code actually reaches, so those go first.
github.com/lambdasec/frame
Good timing for the Apple Silicon side of this. Not All Layers Are Equal covers mixed-precision quantization for weights and KV cache on Macs: which layers are sensitive, and what keeping just those at higher precision buys. xitter.gq/latent_node/status/202…
Happy Dhanteras from the palace arena 🪔 Free browser games, no install: mythology.games/?utm_source=…
OpenRouter's new Auto Code routes each coding-agent turn to a model. Boost in OptiQ Code splits differently: a cloud model leads while your local model on the Mac reads files and runs tests. On 30 mini tasks, 4 solved became 15, for $2.80. mlx-optiq.com/blog/optiq-cod…
AI visibility numbers flip week to week. One question on merageo.com went 2, 0, 2, 1, 0, 2 assistants across six scans. meraGEO's new scoreboard only counts a question as named in 2 of the last 3 scans, and sets it next to your first scan. Paid plans: merageo.com/pricing?utm_sour…
The camera drifts over a river city, settles on the palace's gold domes, and dusk falls with every lamp lit. Switch off raids in Myth City and you can just build and watch. Free to play. mythology.games/city/?utm_so…
Eight riders on one long road, and you can hit the one next to you. This is the snow stretch on a bull in Myth Dash, free in your browser. mythology.games/dash/?utm_so…
On Easy, a bull-drawn chariot starts 8th on a seaside track, works past the elephants and the crowds, and wins the race. Myth Kart is free in the browser, nothing to install. mythology.games/kart/?utm_so…
A low AI visibility score says there's a problem, not where. meraGEO now lists the questions where ChatGPT, Claude or Gemini named competitors and not you. These are sampled answers, not search volume. On paid plans, and the first scan is free. merageo.com/?utm_source=x&ut…
Qwen3.5-4B-3D-Coder now runs on a Mac as a 4-bit OptiQ quant and writes game assets as single-file Three.js HTML. About 26 tokens/s, 4-6 GB peak on a 24 GB M4. This cannon is one of the 11 that rendered out of 18 tries; 3 crashed, 4 looped. mlx-optiq.com/models/qwen3.5…
meraGEO's own scan gives merageo.com 3 out of 100. ChatGPT named it in 1 of 8 category answers, Claude and Gemini in none. Profound, Peec AI and Otterly.AI came up instead. It already has an llms.txt and a sitemap. Rescan on Dec 1: merageo.com/scan/merageo.com…
A miss from a Claude Code session on Opus 5.5 in auto mode. During routine file moves it added rm -f /dev/null to a command, for no reason. Auto mode didn't stop it. macOS refused, so nothing changed, but sudo would have let it through. It flagged the slip itself.
A data breach notice arrived today: an outside party got into a company's internal systems and copied customer data. The cause isn't public. Check your own first: Frame scans code, dependencies, secrets and git history locally, and proves code findings. github.com/lambdasec/frame?u…
The OptiQ 4-bit build of LFM3-1.7B runs on a Mac at 1.3 GB, down from 3.5 GB at bf16, and beats plain 4-bit on IFEval (72.6 vs 66.5) and HumanEval (72.6 vs 64.0). mlx-optiq.com/models/lfm3-1.… xitter.gq/maximelabonne/status/2…
