sebmellen

For those of us with strong verbal processing and weak symbolic/pattern processing, this makes math much more difficult to approach.

Interestingly, discrete math feels the most "verbal" of all the subfields of math I've encountered (I haven't gone very deep). I think this is because notation in discrete math is is somehow closer to compressed prose or logic, whereas other forms of math use notation to fill in for long sequences of symbolic manipulation.

Not sure if that makes sense... I'm curious whether anyone else experiences it that way.

Will be switching my blog to this.

Surprisingly, Kagi is actually just a really good business. Look at my comments, I despise AI spam and call it out wherever possible.

That phrase itself is such an astounding performative contradiction [0].

0: https://en.wikipedia.org/wiki/Performative_contradiction

What brand?

This must be insanely difficult to get right. I was not expecting such an impressive result from a site that looks so vibe coded! The look is underselling how good this is.

They hid it well at the bottom of the page: https://lmstudio.ai/

You've got to grab a screenshot of that.

Wow, $139 B.

That job description feels so far beyond parody that I could scarcely believe it until opening the link! What a world.

Are you using a local model or one of the hosted ones? I don't think a harness can influence the range of outcomes that much.

Wow, I like that a lot more than what it gave me! Shows you how wide the range of outcomes can be.

Yagil, thanks for the credits! Lots of fun to try running this locally. I don't understand the negativity in some of the other comments.

This is one of the better agent harnesses I've seen for inspecting reasoning chains, which is super useful for me. Sometimes reading the reasoning is better for my needs than reading the response. I appreciate how transparent that is in contrast with Claude Code/Codex/etc.

One question — I saw you mentioned negotiating ZDR with your "providers" — are you hosting the models yourselves, or is another entity hosting them? If another entity, which group?

Very neat, this is definitely a more complete prompt than mine!

My testing prompt for these models is by no means objective or repeatable (like the pelican) but it's a nice test of curiosity:

Impress me with a 1 page html file

Result: https://ydaurtg3fdwhq.kimi.page/

Came out looking pretty cool! By contrast, Fable produced a moderately more interesting "live observatory" of the solar system.

the figure is before contractor payouts

Ok… so it’s GMV, not ARR… Can we say eBay hit a gross run rate of $80B?

Yeah, but as you say, this is another LLM rewriting it. The amount of noncontextual information is nauseating and destroys the point of a README (in my humble opinion).

A human might have written a disclaimer like this:

When not using Fable, pxpipe may require additional follow-ups to precisely follow your formatting instructions.

This kind of garbled information dump is very inconsiderate of the reader, and all good writing is considerate of the audience consuming it.

It’s so painful to read the LLM-compressed explanations. I can’t exactly identify what it is, but it’s an immediate tell and literally requires twice the effort to comprehend.

For example:

Honest caveat, visible in the clip: the pxpipe arm answered the count first and needed one follow-up nudge to also print the ledger balance in the requested one-line format; the plain arm followed the format on the first try. Legibility is solved on Fable — single-reply format compliance is the remaining rough edge.

If I reread this four times, I can sort of interpolate what happened, but it’s mostly pointless and confusing information.

In my experience all models do this to an extent, but Claude seems to be the worst at this. GPT 5.5 is a bit more terse but seems to compress more valuable information.

I’ve found unexpected success in using ephemeral NixOS VMs for local development… once you authenticate your agent you can let it run wild without worrying about permissions.

This is really far too broad a brush.

Do most medical students publish useless case studies trying to jockey for residency spots and signal hustle/devotion? No doubt!

But there are a good handful of medical students who are still (surprisingly) in it for the medicine and not the money. And that handful is exceedingly capable; no reason they can’t publish valuable work with the right collaborators and resources.

I suggest we send this fellow to the Monty Python Argument Clinic https://youtu.be/TpQlyUjp3vM.

Very cool concept with AltTab. Have always been looking for something like it!

Generally speaking, every YC company post ~2020 is forced to make pathologically false claims to compete in the (fundraising) market.

Agreed

Just commenting for posterity… if this is what it claims to be, I am not looking forward to how it will empower the people who submit bug bounties to us.

Historically they’ve been people from certain identifiable countries (usually developing/poorer countries) using fuzzers with low-quality results.

Now, those same people use the current-day models to good effect, but they still don’t have a true security edge and oftentimes the reports are minor or duplicative.

I wonder if that’s about to deeply change.

Ok. Claude will not work for this use case because none of the sample data (weirdly blurry ID images) is in the training data.

Great, let me know when those models can run on-server and process/analyze streams of ID images with less than 100ms of latency. You’ll need to make sure you have a massive set of training data including all manner of slightly blurred and slightly distorted ID cards

I would never disgrace you with reading my LLM output unless I explicitly identified it

If I'm interpreting this correctly, GitHub will use their existing actions infrastructure to run versions of the code in isolated worktrees. I think this could be a very beneficial process.

What I've done on my end is created a project where I have a remote Linux workstation. I can create multiple worktrees for each repo in that workstation, and then my agent can push PRs to GitHub and use the actions infrastructure to see if the integration tests that it writes for itself are successful without needing to run those integration tests on the local environment. It's expensive in terms of runner hours, but the automaticity of it is incredible.

The most frustrating thing to me is to receive a 5-paragraph-plus email that was clearly written with some AI that filled in the email with vapid and useless talking points, like "Let me know if you need any other blah blah blah; While there is clearly a need for system improvement, we are working hard to address the underlying and fundamental issue; This is a lesson that it's not just a feature, it's a critical path for our users, etc."

My theory is that people are fundamentally averse to the thought and effort it takes to write a good quality email. Then there’s probably some underlying belief that more volume shows more effort, which people will perceive positively. And finally, there's the worry that if you write the email yourself, you might make some embarrassing wording, grammar, or spelling mistake.