jgilias

Quality matters though. A 10 years warranty as long as you do the prescribed maintenance with hassle free ownership is a major reason why Toyota’s clientele sticks.

It’s not impossible to unseat them. But it’s also far from given.

But… why???

This probably throws a spanner in the wheels there:

> Every piece of reasoning the frontier model does for free has to be rebuilt here as deterministic state.

EDIT: Not to shit on this though. I totally believe that some smart mixture of LLM-reasoning + Jev-style + determinism is going to be pretty amazing.

What harnesses do you use? Ours is basically Claude in a box. There’s some complexity because of that, but the advantage is that it’s very flexible and people who have a bunch of Claude-shaped skills can just basically give those to an agent.

I’m thinking to take a deeper look at Pi. I’m really liking that project.

At $DAYJOB we do something very similar shape-wise. I wonder - it sounds like you have a dedicated agent comms plane? In our case we found that the easiest and most straightforward was to just use our default company chat app directly for this. Because most of the context that the agent workers need to do work is there, but also, it’s just much easier for teams to conceptualise an agent colleague if it just hangs out in their channels.

What do you do here, and how’s it going?

Maybe it does. Need to run evals to see if it does or doesn’t.

Point was - everything in the context window affects the output. Including “silly” things like “it is known AI can do this”. And that has nothing to do with superstition, as the poster above me seemed to imply.

Not OP. That’s not implied at all. The fancy autocomplete produces statistically likely continuations to the source text (the context window). For a problem that’s hard for humans one likely continuation is: “this is hard, can’t do”, even though there’s enough in the training corpus of the LLM to actually do it.

So, it follows that adding “pep talk” into the context window reduces the statistical probability of “no, can’t do” coming out as the answer you get.

These things are neither humans, nor deterministic software.

If it’s 3 days it’s something like 15 minutes, if it’s 3 weeks, that takes a couple hours lol. Seems like there’s some sanity to the estimates after all when you think about it, it’s just the scale it gets wrong due to estimating human time.

A contrarian take - the models aren’t advancing anymore at a pace where each new model would represent a huge capability jump, all being incremental improvements, so the doomsday marketing strategy being invoked since GPT-2 isn’t as effective anymore. Then “pacing” would be a convenient scapegoat to point fingers to when people point out how the new model isn’t _really_ that much better.

“Of course it’s not, we’re pacing!”

I chuckle that this is in Windows XP days. You know, the pinnacle of Windows usability in retrospect.

It does matter, yes. Because it won’t lead to the outcome you want until someone actually gets a painful fine.

In medicine there’s this well known axiom that if an intervention regime would technically work, but is something that the patients tend to not follow through, then the intervention regime doesn’t work.

I’m so tired of every model release being touted as AGI or similar. Since <checks notes> GPT-2.

Cool. I’ve realized though that I don’t really need better models anymore. SOTA is good, I just want them faster/cheaper now.

Volkswagen has a couple too. Like, actual laboratories with people in white coats working in them. You don’t call Volkswagen “a lab” because of that.

Ok, so, what research department did Nvidia grow out of?

_why_ have people started to call companies “labs”?!

Would be cool to get this thing try and run the vending machine experiment.

That’s not surprising. He’s, like, really dense.

Hooked me with the first couple sentences, and I just had to keep going, as I just wanted to know what happens next.

Thanks, a good read!

You have it mixed up. This guy is “the new generation of product people”. The old generation had to be technical enough to be able to grok the systems that they ‘producted’ over. It’s exactly the ‘new generation’ that LLM themselves right into Dunning-Kruger.

For the completion to work, the source text needs to be ‘good’. That’s a basic kernel of how the thing works. Even with a perfect oracle autocomplete if the source text is ‘bullshit’, the output is too.

Or, slightly changing this. The source text needs to speak the correct vocabulary and language to produce a good completion. See the chat where Terry Tao is doing maths with an LLM. There’s _no way in hell_ I could get to his output because I just have no idea, and can’t speak the language.

Same with any field.

We have a product guy on the team who was in a deeply not technical role before AI who is trying to do the “hey Claude, read this Jira ticket, implement” thing.

It doesn’t work for the vast majority of tickets he attempts because he doesn’t have the necessary understanding to even start thinking about if the solution that the autocomplete generates is even remotely workable. And that’s with fancy dev loops and whatnot.

The spacer between the keyboard and the chair still matters in my experience.

Being a European, I first upvoted, then moved on to read the article. I assume I’m not the only one.

And my corner of Europe is relatively not that red-tape-ish. But many places are pretty horrible when it comes to that.

My neighbour has a battery installation because where I live they subsidise both - the panels and the battery. In the event of a power outage he can keep running his house for ~3 days if being somewhat frugal with consumption.

And the battery thing even looks kind of neat.

The more they fiddle with the autocomplete system, the more they move away from the autocomplete faithfully producing the completion I need. The more it makes sense to move to an open weights model not served by them.

Definitely. I can easily cut my individual AI spend by 50% at least. But I won’t, until someone tells me to. For three reasons:

1. Thinking about how to optimize AI spend is one more thing to think about when I have too many things to think about already;

2. If there’s even a minuscule chance that me selecting down the model results in a task getting delayed because for that particular thing the model happened to be “too dumb”, I’m just not going to do that. Because I’m evaluated for what I ship and saying, “oh, not happening this week because I downgraded to Haiku for the exploration phase, and that resulted in a faulty plan, so it slipped a day” just doesn’t cut it.

3. Why should I spend time thinking about costs, and risk being slower because of that, if the next guy doesn’t?

I wonder if there’s path dependence there. It _seems_ like just because you build up lots of “defence” tech that doesn’t necessarily mean you _have_ to use it? Or does it? I wonder what happened with those markets during the Cold War.

What I’m doing is running a kind of a meta harness that uses different models (and underlying harnesses) to work on a problem, or review the solution. Idea is to get to an error rate better than each of the underlying models can provide. Same thing as sensor fusion.

Now, that’s slow and expensive although seems to work quite well (haven’t really evaled this properly, don’t have the time). If inference can be made fast and cheap, multi-model approaches like this would become more viable for more applications.

Oh boy, thanks for sharing this, truly mind blowing. It was chatjimmy.ai