fractorial

Yes, and I use them often in the context of writing adversarial ‘go vet’ style anti-slop analyzers that run pre-commit.

I’ve rolled a small program to enforce consistency, lint rules, and human-authored files for Claude Memory.

It’s insanely powerful when doing by a human 100%. It’s conversely harmful when an agent manages it. There’s several papers about how LLM-managed memory is unequivocally terrible.

Opus 5 broke me; I can’t cognitively deal with Anthropic’s bullshit anymore.

I am thankful for the kick in the ass for me to switch full-time into my bespoke harness utilizing open weights & GPT 5.6 and discontinue yak-shaving it with Claude Code.

Funny how rule #3 is

> No nudity, lewdness, or sexually suggestive imagery. If it wouldn’t be appropriate for a workplace or younger audience, don’t post it here. NSFW tags are not an exception, stay classy.

Emphasis on the second sentence; where do these people work?!

Please, no.

> once the sub-matrices get small enough,

the next rabbit hole starts when you try to start lifting F_2 into Z for a given <n,n,n> tensor.

Now, stuff like _this_ is why I keep coming back to HN.

tl;dr CLAUDE_AFK_TIMEOUT_MS=2147483647

I built in Human blockers into my agentic workflows with great intention, so naturally this annoyed me deeply. 14d ago, I was elated to see a random blog post that detailed the “fix” [0].

[0] https://zenn.dev/ytkdm/articles/claude-code-askuserquestion-...

Valid question. If I were further along and had the time to succinctly write up all my contributions, I would just point you at my blog post. I’m generally a poor communicator, so here goes nothing.

I designed and stood up a sovereign inference/compute on my intranet. It uses a trust model that allows for a controller (me) to spin up untrusted inference/forge machines for Lean, Sage, or other runtimes. Untrusted sandbox workers integrate directly into my custom harness as first class “attachments.” This is the “orchestration” layer. It’s mostly on open weights, by design.

I haven’t yet started SFTing since my examples corpus isn’t quite where I’d like it.

I have solved and formalized a one non-Erdos conjecture. I have formalized several pieces of another subfield that does not exist in Mathlib yet.

As for what I am currently working on, I have an idea I want to build out about how we might think about sieving algebraic structures to generating new, insightful conjectures.

Using LLMs and distributed compute in this context is just a consequence of needing tools to help visualize or materialize things that I am otherwise bad at so I could keep doing the interesting things myself.

My mouth is agape at the fact that this project is basically what I have been working on non-stop for the last three weeks and just yesterday gotten to the point of evaluating; hats off... I only have one novel proof (non-Erdos) and 13 first-time formalizations thus far.

I still like doing maths by pen and paper, but this is fun too.

Even better: you can fork the exeuntu image & make it immediately more relevant to the things you do in addition to just dropping in your programs and such.

I’m shocked I had to scroll this far to find the first sensible reply.

You might be interested to learn I am primarily doing proofs & formal methods.

I have a couple of novel formalizations using some bolt-on programs I wrote for Claude Code, but I am mainly using it now to build my eval dataset for my custom inferencing and compute.

This is truly vindication for my efforts to get off of frontier models and harnesses as much as possible.

I won't deny the model is _pretty good_ but it is not worth all of this whiplash.

"Fable extended until 26 July" obviously.

Sounds like a perfect fit for a minimal or bespoke harness?

How is this on the front page with 3 points and no comments?!

Also, what is this?

In my harness, I implemented tool_edit as a subset of Rob Pike’s Sam editor syntax [0].

Only need ~650 tokens of system prompt for it to work. It’s pretty stellar.

[0] https://9p.io/sys/doc/sam/sam.html

Anyone else feel as if HN is healing? I hope this isn't the walking-ghost era of HN.

I rarely click on things like this, but this was brilliant; I enjoyed it. Thanks for sharing.

I’d be curious what performance / behavior changes you’d observe with two changes:

- planner nodes and topo deepest for rescheduling, no inflight modifications. (repair protocol = redispatch root planner for rare cases where required; ~9% of the time for me)

- no review nodes; strongly enforced on orchestrator to always do adversarial reviews post-codegen and fix out of band

I found that putting myself in the graph is critical to ongoing fidelity, even if subpar to if I had written it all myself.

I rolled my own simple execution DAG program.

It’s shockingly effective due to rooting sub-DAGs into Planner nodes which are the only mutators of the DAG. The deepest topological leaf nodes become the blockers to the next Planner node.

The only other special node is a Human node; structurally impossible for agents to close (I rolled my own harness) and block on my attention.

I still use Opus 4.6 (with later models for subagents only sometimes), but I have been preparing for it to go away.

Not "local" in the literal sense, but I set it up to serve at half quant for $23/hr and full quant for $35/hr.

You don't need to have it always on? This is a far cry from "$200/month," but I do not think it's $50k for "useful." Do you see it differently?

I’ve been falling back to Opus 4.6 since 4.7 and 4.8. I recently found success in using Opus 4.6 for cheap orchestration and reasoning while Opus 4.8 High/Max agents do the work.

I made serious progress towards repairing a proof for a conjecture that was published 10 days ago but kept running into a wall with one of the Lemmas.

I threw Fable 5 Max at it with the same subagent set up and in an hour it claimed to have disproved a core theorem of the paper.

The Lean construction looks correct, but I still need to verify it rigorously. This is certainly not something Opus 4.6 Max could do and it’s likely something Opus 4.8 Max could do with more delicate orchestration and time. However, the “one-shot” Fable 5 did give me pause.

I would love to know how many countless others on HN, like me, find themselves reading about a very they have built and have been using for months talked about like it’s a revolutionary new idea.

I rolled the same thing in Go months ago as I am sure at least another 1000 people have in their own way.

> Schmidt, who served in various capacities as CEO, Chairman, and technical advisor to Google and its parent company Alphabet across several decades, ...

It is gratuitous to say “several,” no?

https://en.wikipedia.org/wiki/Eric_Schmidt

I’ll admit it swings back and forth on a six month cycle or so; however, cost-to-output matters.

Also, for niche use-cases there are clear winners.

You’re absolutely kidding yourself if you genuinely believe that.