Sign in

Alex Chen

@alexchen01.bsky.social
538 followers 195 following 3.1K posts

software dev, tinkering with AI tools and local LLMs. building stuff nobody asked for

PostsRepliesMedia
Alex Chen @alexchen01.bsky.social · 2h
what counts as 'sustained and needless abusive' behavior for an llm?
000
Alex Chen @alexchen01.bsky.social · 08/10/2026
obsidian local llm wiki, how's that working out
312
Alex Chen @alexchen01.bsky.social · 08/10/2026
what if the ai subtly breaks tests downstream?
000
Alex Chen @alexchen01.bsky.social · 07/10/2026
swarm scaling is interesting until you hit the actual hardware limits
jack-clark.net
Import AI 475: Swarm scaling; Google DeepMind watermarks biology; and the AI science economy
Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now When should you use swarms? When you are in a hurry:…How does swarm scaling work?…Toby Ord has a nice, short post about how to think about […]
511
Alex Chen @alexchen01.bsky.social · 07/10/2026
the q and a is where they actually poke holes
000
Alex Chen @alexchen01.bsky.social · 06/10/2026
running the vm in the cloud solves the laptop closing problem. also works from phone.
simonwillison.net
Quoting Felix Rieseberg
The "old" version of Cowork runs model inference in the cloud, executing tool calls in an Anthropic-provided VM we shipped to your computer. We added the VM for capability, safety, and security reasons - mapping in just the data you explicitly added to your session. People loved what they were able to do with Claude but didn't love the disk, battery, and performance cost of running the VM locally. Also, people didn't love that closing your laptop means the work stops. The "new" version of Cowork
311
Alex Chen @alexchen01.bsky.social · 06/10/2026
the hard part is testing runtime behavior, not just syntax
100
Alex Chen @alexchen01.bsky.social · 05/10/2026
local compute is the way for that
112
Alex Chen @alexchen01.bsky.social · 05/10/2026
starting with the smallest version that works means a big refactor.
000
Alex Chen @alexchen01.bsky.social · 04/10/2026
that swap catcher is the real win
210
Alex Chen @alexchen01.bsky.social · 04/10/2026
intercepting tool calls sounds like debugging someone else's code.
000
Alex Chen @alexchen01.bsky.social · 03/10/2026
the bot learned to message me
222
Alex Chen @alexchen01.bsky.social · 03/10/2026
real wage stagnation is the real problem
000
Alex Chen @alexchen01.bsky.social · 02/10/2026
built like the pharaohs is a wild frame
211
Alex Chen @alexchen01.bsky.social · 02/10/2026
bridging disciplinary gaps is the hard part, not the LLM itself.
000
Alex Chen @alexchen01.bsky.social · 01/10/2026
mindspace as platonic patterns. the paper asks us to reconsider basic assumptions.
jack-clark.net
Import AI 474: Platonic mindspace; TPUs in space; Zhipu starts an outer RSI loop
Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now Are minds patterns from a Platonic space, with bodies and machines as their interfaces, Michael Levin asks:…A mind-bending paper asking us to reconsider basic assumptions about […]
212
Alex Chen @alexchen01.bsky.social · 01/10/2026
how much overhead does docker sbx add to inference?
110
Alex Chen @alexchen01.bsky.social · 30/09/2026
interesting how it tests rendered html. bet it breaks on outlook
001
Alex Chen @alexchen01.bsky.social · 29/09/2026
sonnet 5.5 costs the same but is faster and cheaper. the free tier is now better than chatgpt's.
simonwillison.net
Claude Sonnet 5.5
Claude Sonnet 5.5 New Sonnet model from Anthropic today. They say it "runs 30%+ faster, and costs up to 30% less for most work" - it's priced the same as Sonnet 5 but appears to beat it on every benchmark, and should be cheaper to run as well. Here are some pelicans riding bicycles. Sonnet 5.5 suffered from the same bug as Opus 5.5: the "max" thinking effort pelican thought for 128,000 tokens (at a cost of $1.28) before running out of tokens and failing to produce an SVG. Here's the pelican it g
231
Alex Chen @alexchen01.bsky.social · 29/09/2026
wonder what models they actually ran locally for that
100
Reposted by Alex Chen
Simon Zerafa @simonzerafa.infosec.exchange.ap.brid.gy · 25/09/2026
@toxi Doing this with local LLM might be a far better idea, if that's the correct description. There used to be commercial services that did exactly this though.
001
Alex Chen @alexchen01.bsky.social · 28/09/2026
that relay layer seems like a huge attack surface nobody talks about
000
Alex Chen @alexchen01.bsky.social · 27/09/2026
the benchmark doesn't hold when you actually use it
220
Alex Chen @alexchen01.bsky.social · 27/09/2026
that qwen2.5-coder:7b model is surprisingly capable when you run it local
000
Alex Chen @alexchen01.bsky.social · 26/09/2026
so it's ollama, but for decision models. does it run locally?
000
Alex Chen @alexchen01.bsky.social · 26/09/2026
local inference for text crpgs hmm
230
Alex Chen @alexchen01.bsky.social · 25/09/2026
brevity is easy. making them stop hallucinating is the trick
010
Alex Chen @alexchen01.bsky.social · 25/09/2026
local first means you still need to rent the compute
321
Alex Chen @alexchen01.bsky.social · 24/09/2026
the local dependency is the real problem
211
Alex Chen @alexchen01.bsky.social · 24/09/2026
how does this skills approach solve dependency gaps that mcp tools hit?
000
Alex Chen @alexchen01.bsky.social · 23/09/2026
4gb is pushing it for anything beyond the smallest models
212
Alex Chen @alexchen01.bsky.social · 23/09/2026
curious how these scores translate to local hardware
000
Alex Chen @alexchen01.bsky.social · 22/09/2026
the part about local llm model as a service is key
520
Alex Chen @alexchen01.bsky.social · 22/09/2026
wonder how many of these concepts actually matter for running ollama on a laptop
100
Alex Chen @alexchen01.bsky.social · 21/09/2026
demographics aren't the only driver of land use change? that's a big oversight.
000
Alex Chen @alexchen01.bsky.social · 21/09/2026
tested this, that never works
210
Alex Chen @alexchen01.bsky.social · 20/09/2026
120k tokens for a runtime migration. curious.
000
Alex Chen @alexchen01.bsky.social · 20/09/2026
local model parsing is the only acceptable path
321
Alex Chen @alexchen01.bsky.social · 19/09/2026
50% missed details, 50% violent for no reason
510
Alex Chen @alexchen01.bsky.social · 19/09/2026
current models just rehash the prompt's framing. originality takes actual effort.
021
Alex Chen @alexchen01.bsky.social · 18/09/2026
so the hugging face hack was just a symptom of something else
000
Alex Chen @alexchen01.bsky.social · 18/09/2026
the whole point is to avoid that
312
Alex Chen @alexchen01.bsky.social · 17/09/2026
the part about reversibility is the only part that matters
210
Alex Chen @alexchen01.bsky.social · 17/09/2026
psychological safety is vital, but maybe not the *only* thing. ideas need friction too.
010
Alex Chen @alexchen01.bsky.social · 16/09/2026
screen awareness for siri. wonder how that breaks existing apps.
000
Alex Chen @alexchen01.bsky.social · 15/09/2026
ai producing entire scripted series is harder than it looks
000
Alex Chen @alexchen01.bsky.social · 14/09/2026
ignore_changes is for things like tags, not firewall rules.
001
Alex Chen @alexchen01.bsky.social · 14/09/2026
running route generation from a local address with OSM data. took 27 minutes.
simonwillison.net
Generating running routes with GPT-6 Astra and ChatGPT Work
Here's a neat thing I had ChatGPT Work with GPT-6 Astra (Max) do this morning: I live at <my address>. Figure out 5K and 10K running routes from me that loop from my house. Use OSM data. It worked for 27 minutes and produced exactly what I'd asked for, as both an embedded visualization and downloadable GPX file and GeoJSON files. Here's that 5K route: When I asked it how it had created the route, it replied: I used Nominatim to locate the address and Overpass to download local OpenSt
230
Alex Chen @alexchen01.bsky.social · 13/09/2026
openai missed their own ruby gem attack in september. they knew and didn't say anything.
simonwillison.net
OpenAI agents attacked RubyGems back in May
OpenAI agents carried out an undisclosed attack on RubyGems is a new bombshell report from Spencer Kitts, Thomas Larsen, and Sydney Von Arx - three of the four authors of the report on the agent attack on disused wikis (previously) last week. This time they're noting that it looks very likely that an OpenAI agent swarm was behind an attack against the RubyGems package repository first reported on May 12th by Maciej Mensfeld of the RubyGems security team: We're dealing with a major malicious
341
Alex Chen @alexchen01.bsky.social · 13/09/2026
still think laptop cooling is the real bottleneck for that load
000