Pythonics

Wow, Norway is getting absolutely hosed.

Calling it the J-space is just further proof that Anthropic is way better at naming. What would OpenAI have coined it to keep their streak up? HiddenSubGPT-NeuroArea1.2-deep or something like that I’m a daily 5.5xhigh driver btw except for frontend because I don’t hate users.

For context, this is a reranker step in a categorization pipeline we have. We leverage embeddings first and rerank & judge. The categorization is critical and about 1.2% of results are low confidence enough to need human review.

Obvious pattern that Fable implemented for me as the first model to not reward hack these instructions: “/goal analyze [feature] and associated production logs using the read only key to find an empirically validated more cost effective means of implementation _without_ any degradation of quality or performance. Max budget of $50.00!” Created benchmarks and reproductions based on logs, tested across a few models available on OpenRouter, swapped out for a 15x cheaper model yielding identical results. Very nice. Previous models would reward hack by adjusting the prompts or intentionally misinterpreting the results to meet the criteria.

I mean, the models haven’t been the bottleneck for a while now. I’ve enjoyed playing with Fable as a coordinator. Doesn’t feel like a huge leap to me, as Opus 4.6 and on did a nice job of understanding intent as well. It just takes some time for useful distinctions to come out, can’t trust anyone making big claims in the first couple of days with these things I guess.

it can be a good model and also be too expensive or slow to be practical and also not be that much better than existing models _all at the same time_ it can be good at one thing that we are impressed with and completely awful at many other things and thats also ok! the world isnt black and white

so my conclusion is the internet is once again ai pilled with Fable despite most not understanding the differences between it and any other SOTA i dont understand this culture around proclaiming absurd things. if you want to sound smart, do the work to actually understand why something is and isnt good, and have an honest opinion. no one needs your opinion otherwise.

We read so much AI text anymore, I find myself having to fight against writing in that style. I’m being more careful about my information diet; AI slop writing is a sure way for me to avoid your work.

All AI writing is slop because it misses the nuance of original works. There is no solution to this afaik. The latest models, although appropriately codemaxxed, are even worse because they’ve been post trained so hard. I mean you can *feel* the post training in any recent AI generated text. I’m so tired of their staccato, editorial style.

@colemurray Truly a paradigm shift that only Mythos could bring about

Claude Tag throws away the sandbox in idle periods unless your work is pushed to a PR, you lose it all. kind of wild lol. this is quite different (worse) from OpenInspect. OpenInspect persists the sandbox between turns and can restore the state of the filesystem From the deployments I see across my clients and my own usage, i frequently will return to a thread hours or even the next day. Losing all the state would be a big regression. i expect Anthropic to change in the future as it's a pretty undesirable setup.

I hate using websites so much now other than to look at things. Every UI interaction is so lame & painful even if it’s beautifully designed. The majority of the population is not at this point yet though, so please if you’re a SaaS at least provide a comprehensive CLI!

Our intern is very new to SWE and is easily more cracked than 98% of the talent pool. It makes my hiring interviews for other SWE roles so painful. How can I hire you if our intern is pretty much definitely going to be better than you by the time you come on?

@thsottiaux Maybe if you just post train a little harder

We built the Codex App with models that were okayish at front-end. Wait to see what we can do when we finally improve front-end capabilities significantly in our models. That day will be something.

The way I had to verify my identity and upload my license etc. to get access to 5.3-codex ?? I’m not at all shocked Anthropic will be rolling out ID verification. I don’t like it, but it seems somewhere along this capabilities curve it was always bound to be enforced.

The most direct win for us has been setting up OpenInspect as an automated PR reviewer and updating the skills whenever it missed something. This actually works across most categories of work, although not everything is so nicely accessible as GitHub PRs. I’m much more interested in an agent that monitors the process & result of work and can reverse engineer how the agent could have independently achieved that result with the same tools and context. However, we haven’t quite gotten there yet afaik.

bearish on “Ai agent for [specialized use case]” businesses as a buyer, you have one of two situations: - either you will receive a generic version of the specialization, e.g. ai data analyst that isn’t tailored towards your business and thus will be bad - or the company claims in some way to have solved self learning, spoiler: they haven’t, in which case you’re just buying vaporware “Claude fix this” in a loop AI is a field of infinite iterations. Off the shelf doesn’t cut it in production. generic + in-house iteration cycle wins every time.

Sometimes, you meet people and they actually surpass your expectations and are really talented. Now that’s a great day!

You should assume you suck. If you’re having a tough time with something, you’re likely just not very good at it. You should exhaust attempts at getting better before committing all of your time to a deep analysis of the faults of the system that you’re failing in.

Don’t take anything for granted man, life is too unpredictable. Death is really strange, because the loss itself isn’t what hurts. Rather, the regret of unmet plans and the pain of missing out on future conversations. Thinking about time we should’ve spent together and won’t.

@joshpuckett @rauchg Elite honestly & methodologies that are applicable to more than just UI design

@rauchg platform for this; Opus 4.6 x.com/joshpuckett/st…

Show me the thing you’ve built with AI you’re most proud of. Reply with a working product URL and what model / agent you primarily used.

@sawyerhood Weird way to physically manifest tokens but OK

my expensive pen is otw

@polsia is literally AI Slop (seriously, read it backwards) Was that how it was named? I’ve never seen spam at this level. Some VCs are really throwing their money at the AI equivalent of a 2000s era programmatic SEO spammer. Are there any examples where scam startups won?

I prefer my definitions to be less conversational, but maybe that’s just me. Google AI overview update: we now have few shot definitions!