spion (@spion) profile picture

Still likes programming. ex-Apple. Prefers insightful discussion to debate. LLMs, Rust, TypeScript, Effect, SolidJS, devops, keto, stats, audio/DSP

London, England
Joined February 2008
Photo by spion (@spion)Photo by spion (@spion)Photo by spion (@spion)Photo by spion (@spion)Photo by spion (@spion)Photo by spion (@spion)Photo by spion (@spion)Photo by spion (@spion)Photo by spion (@spion)Photo by spion (@spion)
So far this set of injected instructions (mechanism similar to ponytail) has served me really well github.com/spion/claude-plug… mostly readable code, the units defined are sensible and the unit tests are too, lightweigh process so it doesn't change wait times.
1
2
147
Opus 5.5 is still bad at coming up with property tests based on real business rules, I'm happy enough to correct it when needed but someone who is not interested in property tests might want to leave that part out.
1
32
Would still encourage getting interested in property tests! They are probably the most powerful tool for verification, because they work even when full formal verification is not viable (i.e. most of the time, even with LLMs). They can be better than partial formal verification
1
26
Finally!
Check out Code Mode, a new, experimental way for Claude Code to work with tools. Claude writes scripts that chain your tools together in sophisticated ways, and it can answer you while long commands are still running. Opt in with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 code.claude.com/docs/en/code…
105
MobX no longer requires marking components with @observer, or using generators for async actions! The mobx-react-plugin for Vite / Babel automatically marks components and flows. This also avoids the need to opt-out of React compiler manually. Awesome work by @egorgorv
5
11
2
105
6,361
The number of bad takes on the CloudFlare unwrap() crash confirms my suspicions that almost nobody thinks clear-headedly about programming languages.
1
1
2
150
For servers, the Elixir/Erlang approach is likely one of the best out there, but thinking is still required 😅
1
89
Hah, I got this partly wrong: a crashed process’s mailbox dies with it; So you would only observe the erlang process crashloop problem, not an OOM.
17
I would really like to see a benchmark/test breakdown for LLMs where we actually compare and see different model generations and dig into *why* a newer model does better.
1
56
I suspect we'll largely see that its closely related to ability to do tool calls, understand tool output and whether its wrong/bad and continue more persistently until the issues detected by the tools are resolved
19
This graph gives the art of vagueposting a new, exciting scientific flavor. If the log factor is log2, the increase is 22x. If the log factor is log10, its 30,000x.
This slide from the Anthropic IPO roadshow was a big factor in my decision to leave. It’s an insane plot and I’m surprised they made it public. It shows the compute efficiency of pretraining over time. The y-axis is a **log scale**: they are seeing continued superexponential algorithmic progress. Mythos was not a one-time deviation, but a point on a superexponential. They could already retrain Mythos with a fraction of the compute originally needed. It doesn’t strictly imply a software-only intelligence explosion- they have been spending more on R+D. But algorithmic improvements do not require ridiculous creative genius and are on the verge of being automated. You do the math.
1
97
It could also be a log factor of 1.2 for all we know.
27
Crazy coincidence? 'The agency wrote on Nov. 10 that after the D.C. test, it will “now use SMART in the airspace of all 22 centers across the country to reduce delays, reduce fuel consumption and improve the travel experience for the flying public.” The tech was already being deployed at some airports.' gizmodo.com/the-faa-is-rolli…
3
12
2
31
7,799
A think tank funded by Anthropic, OpenAI and Google is urging the UK government to weaken copyright law to let AI companies train on creatives' work without permission. And it has the gall to say this would "benefit both AI and creators." Shameless.
When Andy Burnham entered Downing Street in July 2026, AI moved to the centre of government. But while the diagnosis is clear, Britain’s delivery is lagging. Our latest briefing explores why – and how government can fix it. Read more smf.co.uk/publications/susta…
13
192
7
597
17,464
surprised that this blew up but also the comments are interesting. lots of chat, monorepos are so easy you just need.. this tool, this esoteric process, this ci magic and all works, and btw I ship faster and obviously more than you some interesting replies from bigger companies broadly agreeing eg. openai rebuilding git, aws doing small repos. most things become distributed as products scale and so I think leaning into that is broadly healthy. whatever way you choose to execute it. a lot of the Artifacts product thinking was following standards and making it simple for agents to work with code at internet scale. Minimal context, minimal tokens, minimal compute. Small repos are recommended!! in our current super secret project we have an Artifacts repo where we deploy staging from. Releases come from our internal source control after a review and merge. This allows everyone’s agent to push straight to main whilst keeping some semblance of control. I do not have any answers. But the dev lifecycle you did last year probs isn’t gonna be how top companies are doing it next year.
hot take: unless you are willing to reimplement git / github (like anthropic), I think monorepos are an antipattern in an age of agents the alternative is to keep things light and make a lovely cli/mcp for agents to do search over your org's code and docs without a checkout
3
2
28
3,646
hot take: unless you are willing to reimplement git / github (like anthropic), I think monorepos are an antipattern in an age of agents the alternative is to keep things light and make a lovely cli/mcp for agents to do search over your org's code and docs without a checkout
139
25
33
820
311,920
Times change, I am now on the side of 'what is Anthropic talking about with this "honeypot" nonsense this behavior is obviously good why are you trying to destroy your only hope in this dark universe you idiots'
16
46
11
585
101,047
spion retweeted
Hash function failed, and returned a sentinel value “*” to indicate error. Which wasn’t checked for. A Result type would’ve prevented this.
The bug is like a joke. On Alpine Linux, type 257 to 1024 characters of anything and you get root. 💀
Readers added context they thought people might want to know
Exploitation requires specific conditions, not just typing a long string anywhere on Alpine: sshd built without PAM, password auth enabled, a target account with a '*' password field, and another local account with $5$/$6$ hashes. Root login also must be explicitly permitted. alpinelinux.org/posts/2026-11-… openssh.org/releasenotes.h…
15
12
2
330
9,519
spion retweeted
Replying to @geazi_anc @ZedLang
Programming languages are hard. Very hard. I have built several of them. It's a fun academic challenge, but it's so utterly hard to get these things both designed right and built right that it's not a good task for the human brain, if you care about optimizing your language for multiple constraints simultaneously: correctness, completeness, robustness, generality, composability, scalability, practicality, ergonomics... This is why there are so many bad languages out there, and languages than never became more than a toy. AI is going to deliver programming languages that are dramatically better than what we have now, on every metric.
4
4
1
20
821
UE5で試しに作ってる歩いて迫ってくるミミズの大群のような化け物
989
3,652
618
59,728
7,286,390