FuckButtons

I know accountants, most of the stuff they deal with is getting stuff to make sense when your dealing with an unreliable narrator dealing with push back from their employers and then attempting to synthesize a coherent narrative over a large corpus of heterogeneous internal documents, controls, spreadsheets and legal agreements. LLM’s as they exist now are yes men, they are not good at making coherent fact patterns, especially when someone pushes back on them or gives them a red herring or mis-represents something. There’s a long list of things that need to improve for them to gain traction in that industry.

I must be writing very different software than you, I keep opus on a tight leash and it still comes to the strangest conclusions.

If you assume everyone is making 100k it only takes 20 people in a meeting for it to cost 1k.

I think the point I’m trying to make, albeit very bluntly is people often make the banal point that kids these days are too sheltered because they turned out fine. But they’re actually just observing survivorship bias.

dying of old age, what luxury, I’m more worried about starving to death in my mid 50s.

junior dev hiring is down 60%, that’s not just a post pandemic correction.

Counterpoint, my parents didn’t shelter me from shit and my life went off the rails at 12.

This is a partial realization of the idea, but, for a long running agent the proportion of noise increases linearly with the session length, unless you take an appropriately large machete to the problem you’re still going to wind up with sub optimal results.

A 3 bit quant will run on a 128gb MacBook Pro, it works pretty well.

So use a block based cache and tune the block size to maximize the hit rate? This isn’t rocket science.

Yeah, the fact that we have treated context as immutable baffles me, it’s not like humans working memory keeps a perfect history of everything they’ve done over the last hour, it shouldn’t be that complicated to train a secondary model that just runs online compaction, eg: it runs a tool call, the model determines what’s Germaine to the conversion and prunes the rest, or some task gets completed, ok just leave a stub in the context that says completed x, with a tool available to see the details of x if it becomes relevant again.

Bigger model wins as long as the quantization was done properly.

Not entirely true, it’s random access within the relevant subset of experts and since concepts are clustered you actually have a much higher probability of repeatedly accessing the same subset of experts more frequently.

Did he stutter?

Weimar Germany was very socially liberal, homosexuality was socially accepted, legal rights for women were the same as for men, and all of that definitely went away quite quickly.

Today’s llms have had a tonne of deep rl using git histories from more software projects than you’ve ever even heard of, given the latency of a response I doubt there’s any intermediate preprocessing, it’s just what the model has been trained to do.

That’s because it’s superstition.

Unless someone can come up with some kind of rigorous statistics on what the effect of this kind of priming is it seems no better than claiming that sacrificing your first born will please the sun god into giving us a bountiful harvest next year.

Sure, maybe this supposed deity really is this insecure and needs a jolly good pep talk every time he wakes up. or maybe you’re just suffering from magical thinking that your incantations had any effect on the random variable word machine.

The thing is, you could actually prove it, it’s an optimization problem, you have a model, you can generate the statistics, but no one as far as I can tell has been terribly forthcoming with that , either because those that have tried have decided to try to keep their magic spells secret, or because it doesn’t really work.

If it did work, well, the oldest trick in computer science is writing compilers, i suppose we will just have to write an English to pedantry compiler.

> I think there are finely-tuned social algorithms that we innately follow.

That would explain why I can’t do small talk, those are not innate to everyone.

I mean, you can see llms shooting themselves in the foot with c++ right now, just ask them to write it.

They’re using react, they are very opaque, they don’t want you to use any other mechanism to interact with their model. They haven’t left people a lot of room to trust them.

The arrow of time leads towards complexity. There is no reason to assume anything otherwise.

You can get pretty descent results (from past experiments I’ve run) of doing something similar to PyTorch et al, ie, make the compute graph (or in this case, the wiring graph for the synth) in Python then have the real time stuff all inside a compiled extension.

You’re pre-supposing that we can actually afford to just keep throwing more compute at the problem.

Moores law is long dead, leading edge nodes are getting ever more expensive, the most recent generation of tensor silicon is not significantly better in terms of flops/watt over the previous generation.

Given that model performance has consistently trended log linear with compute thrown at the problem, there must be a point at which it is no longer economically viable to throw more flops at the problem.

It’s expensive because superconducting magnets cryogenically cooled in liquid helium are expensive to build and run. No amount of demand signal will overcome that reality.

I have a suspicion that the reason they’re using react, and the reason they’re trying to keep people using Claude code instead of other tools is because it gives them additional data to train on.

I was in the same boat, but you’d be surprised by the number of projects that have embedded lua. Zfs, nginx, redis, haproxy.

I don’t want npm anywhere near my tooling thanks.

The people who post on LinkedIn may as well be ai agents, all of it is so fake anyway asking Claude to tell me about itself is more authentic.

Strange, because I could not for the life of me get Gemini 3 to follow my instructions the other day to work through an example with a table, Claude got it first try.

“But I will lose my X" Some may object that they cannot have a clean desk because they'll lose things. They fear cleaning up…

I intensely dislike the authors smug self satisfied sense of superiority.