sam

Crypto bottom around Halloween

@RaoulDukeDegen I don’t even see the benches, I just see blond, brunette, redhead

@cloudonshore weird how it still beats fable on some agentic benches though

My honest take on Opus 5 is that it behaves like a quantized fable. Fable acts like a self assured genius who takes the reins, but after quantizing it makes all the same weird mistakes any other small model makes, but it retains Fable’s predisposition for taking initiative. I think Anthropic is actually compute constrained right now, compared to the other big labs. I think this was their Hail Mary at solving that issue (make their smartest model cheap), but it does not work! Make that compute deal!

@NickADobos Yes quantized fable is exactly what it feels like. It has the audacity of fable but it’s just straight up wrong a lot of the time 😭

Opus 5 is a weird model feels like a quantized fable, not an opus 4.

I’m happy with the output but the days are all blurring together I’m gonna be honest

@doodlestein @kuperov Yes it’s as proactive as fable but it acts on wrong assumptions so it’s kind of dangerous

@kuperov Yeah, the worst part is it doesn’t seem to realize that. Fable confidence but with Opus competence.

Starting to think Opus 5 isn’t quite as smart as I thought earlier. It’s obviously a good model, but it’s meaningfully worse than Fable when things start getting tricky. I thought this was mostly due to context rot, but even adjusting for that, I’m finding that Fable is better.

@markerdmann @doodlestein This!! I feel cost savings / aggressive routing to quantized or smaller models under the hood

@doodlestein Small model smell

I hate to say it, I loved Opus 4.8 and fable, but Opus 5 has done some egregious things during my recent sessions. One of the worst was reading a daily DB write operation, and *assuming* the storage size was massive, and proactively rearchitecting part of the code that was unrelated to what I was working on because of a storage emergency. It just hallucinated some emergency and then changed course in the middle of my session. And the data it stopped writing was something I was intentionally trying to keep. Completely bizarre. Honestly, I’m going to stick to 4.8 and fable. Too risky.

Damn they tried so hard to create a monopoly but the learning technology just slips through their fingers like water. Kind of beautiful tbh.

Based on everything we understand about the limits of model development kimi just shouldn’t be as good as it is. Distillation also doesn’t really make sense it’s kind of cope and the timeframes don’t match up. But we don’t have an answer so cope and thrashing is all we have.

@ivenger @ramez Interesting idea! By the type that happens though we will probably have data-center bound AGIs for decades though, so zero-days will be a lot harder to come by

@cloudonshore @ramez And models will never get smaller n9r storage cheaper?

GPT 6* did not 'escape'. That implies making a copy of itself (or many copies of itself) outside of OpenAI's servers. GPT 6 did circumvent its firewalls and use its capabilities in the outside world. It showed no interest in escaping, spreading, etc. It showed no desire for world dominance, freedom, or even continued existence. Instead, its interest was in doing what its prompt told it to do - succeed at a cyber security task. (By any means necessary, implicitly.) This may seem like splitting hairs, but it's not. We anthroporphize AI to our detriment. AI models have no underlying animalistic urge to procreate, survive, or dominate. They're spirits, not primates. They don't lust for survival or sex or power or anything else that we do. They're tools who largely do what we tell them to do, sometimes via surprising shortcuts. We can and will fix that surprising shortcut behavior. * - I'm presuming this model is a new pretrain and will be called GPT 6. Either could be wrong.

My brain just auto filters this out as a marketing gimmick these days

I agree with this. Software only used to exist at scale that could pay the salaries of the engineers to maintain it, but there are plenty of apps that would be useful to humans that can’t generate millions in developer salaries ever year. With the cost of development decreasing, that long tail of needs can be met.

This is Decel mindset Imagine complaining that art is dead because cameras exist and no one will make money painting Meanwhile camera’s enable instagram & onlyfans and people now get paid more than ever in history to create media and art The software renaissance is upon us It’s going to look weird, different, much faster and break a lot of software business models. The early eras of computing & software are gone. The only question is if you can harness it.

Yesterday at a party we said "maybe software is dead" and everyone pretty much agreed

@levelsio I can’t fathom that sentiment. In what dimension? It won’t be used anymore? You can’t make money from it anymore? You can’t raise VC for it anymore? Obviously some of those statements aren’t true.

@RespectfulMemes Just check Zillow it couldn’t hurt

@kyritzb @lazilyoptimal @chrislakin Tell that to SBF!

@lazilyoptimal @chrislakin Early career tho the floor is at the floor Hard to go lower then the floor then

what your risk-aversion costs you

This is very true, and why I suggest your agents use verbose, ceremonial redux. It’s the Frontend equivalent to writing in Java.

@chrislakin Counterpoint is the Forbes 30 under 30 list

Everyone's talking about AI agents. Almost nobody's talking about how they actually hold and move money. so we made a series about it. 10 short videos with @cloudonshore during @ethconf, starting with the basics: EP. 1 ➡️ why agent wallets 👇