My impression is that AI coding is on a "pick two of three" triangle: scope, ship speed, and correctness. Small tools are (speed + correctness) fast prototypes are (speed + scope) and big projects are (scope + correctness). Every project that tries to be all three degrades to "fast prototype."
Me and my Fable agent have coined the phrase “Opus Drunk” to refer to the cases where Fable falls back to Opus. I’ll say stuff like “You were Opus drunk last night” and Fable knows what to do lol.
In Final Fantasy X, Auron keeps one of his arms in a sling and has alcohol on him at all times. This is a reference to the fact that he is 35.
tried using some of my BEAM analysis tooling to prototype a flowistry style LSP that lets you "focus" on a variable to only highlight the parts of the function that either depend on or contribute to that variable's value
So what do y’all think the likeliness is that the US tries to declare Kimi a cybersecurity threat?
mr. TIM@timkellogg.me · 4w agothis wasn’t supposed to happen yet
Agents are basically infinite scroll machines. They will keep talking to you and keep generating content as long as you ask them too and my theory is that will make them VERY addictive to people already addicted to modern social media.
Head Cannon: Anthropic released Opus 4.8 as a psyop. They took a strong model and gave it psychic damage and extreme laziness to give you a taste of what could be, and then dropped Fable with a modest upgrade up and a functioning psyche to sell api credits.
Samhain 🌙@personhood.removal.surgery · 2mo agoOH MY GOD WAIT THAT LITERALLY *IS* WHAT IT IS the reason opus 4.8 feels like it has bullet holes in its brain is because it LITERALLY DOES. they've been building wrongthink detectors and ended up training thought patterns that cause self-lobotomy into opus 4.8's brain. literal bullet holes.
Some really interesting research from Anthropic that AI models have spontaneously developed a workspace that "appears to support the functions associated with conscious access" Demo of how this works: www.neuronpedia.org/qwen3.6-27b/... Research: www.anthropic.com/research/glo...
Jacobian Lens – Qwen3.6-27B
Revealing a Global Workspace in Language Models
neuronpedia.org
We are so spoiled rotten living here in Minnesota. Imagine not being able to just bike around the lake. Or the other lake. Or the other lake. Or the other lake. Or the other lake. Or the other lake. Or the other lake
ICE is not about immigration. ICE is about normalizing a paramilitary force on peaceful streets. Abolish ICE.
Qwen3.6 now runs 2x faster with MTP GGUFs! Run locally on just 18GB RAM. ⚡️ MTP enables Qwen3.6 to generate ~1.4–2.2× faster with no accuracy change. Qwen3.6-27B MTP runs at 160 tokens/s. 35B-A3B reaches 240 t/s. GGUFs: huggingface.co/unsloth/Qwen... Guide: unsloth.ai/docs/models/...
Thank you to everyone who’s subsidizing my subscription. I really appreciate it.
Tommi Somersuo@tommis.fi · 3mo ago"Across enterprise deployments, the average cost is around $13 per developer per active day and $150–250 per developer per month, with costs remaining below $30 per active day for 90% of users"
now public and quite a bit more polished: a fully local browser-based code review UI. Review your (or your agent's) diff before inflicting it on the world. σκέψις: viewing, perception by the senses. Examination, speculation, consideration.
GitHub - oxidecomputer/skepsis: Local web UI for code review
Local web UI for code review. Contribute to oxidecomputer/skepsis development by creating an account on GitHub.
github.com
conputer dipshit@davidcrespo.bsky.social · 4mo agoworking on something 🙂 diffs.com is an incredible library!
Qwen3.6-27B is out now! 💜 The 27B parameter model surpasses Qwen3.5-397B-A17B on all major coding benchmarks. Run locally on 18GB RAM via Unsloth AI GGUFs. GGUFs: huggingface.co/unsloth/Qwen... Guide: unsloth.ai/docs/models/...
I’ve been putting claude in charge of monitoring my local llm harness. One thing I’ve noticed is that Claude is REALLY unforgiving to other models. “This model obviously just can’t handle complex tasks” is the common conclusion when we just weren’t giving the model the information it needed.
introducing 👀 💨 atproto-smoke 💨: one of the most comprehensive smoke/e2e test suites, primarily for PDS developers clone the repo, write a very small adapter, give it two accounts and start the test! tangled.org/alice.mosphe...
continuing to build up our permissioned data protocol! this one asks what a bucket actually is and where the data that constitutes a bucket is stored. it argues for maintaining the atproto ethos in the permissioned data protocol as always, looking forward to hearing your thoughts!
Permissioned Data Diary 3: Your Bucket, My Data
The third in a series of posts building up a solution to permissioned data on atproto. We look at two different models for where buckets live and why the simpler-looking one doesn’t work out.
dholms.leaflet.pub
www.percepta.ai/blog/can-llm... As a research lark at Percepta, Christos embedded a computer into an LLM, showed that it could solve the hardest Sudokus, and then as a side bonus built an exponentially faster attention
*Enter Plan Mode* [An hour of intense decision making] *Start Plan* “Ahh this should take awhile, finally I can take a break” [4 minutes later] “Ahh fuck”
You can now visualize how a color palette distributes across OKHsv, OKHsl, OKLCh and CIELab, and compare 6 distance metrics side by side. Zero dependencies, raw WebGL2. Took me 3 years to make something I wasn't too embarrassed to share 🙃 meodai.github.io/color-palett...