Matt “Friend of the pod” Rogish 🇺🇸

Remember when the government turned off Fable for about 3 weeks? That feels like a lifetime ago

Reminder: infra matters, even more now with AI. And in the past, maybe appdevs could get away with poor security practices because it was an outrun the bear strategy: so many other people were bad so what were the odds you’d get hit? But now there are infinite bears on infinite keyboards. Good luck

This week of vulnerability hell is a prime example of why “model safeguards” are anything but. If I cannot use Claude to inspect my code for security holes but the bad guys can use an unrestricted model to find them, there’s no way I can keep up. Anthropic needs to seriously rethink their stance here.

Aint no way Fable is being released before all the resets expire

I used to have Fable 5.1 MAX (or XHIGH) plan everything but it's just too slow and expensive. I can run 30 concurrent 5.5 Opus XHIGH for the price of 1 Fable, and it's all super fast. I then just need one Fable to clean up the mess.

daaang Claude Desktop is pretty rad now

omg did Claude Desktop lose the ability to sort by NAME? AGAIN? Man, maybe their next model isn't as smart as we think it is

where's my fable 5.5 yo

I have been A/B testing Opus 5.5 XHIGH vs Fable 5.1 XHIGH/MAX and so far I'm liking what I'm seeing with Opus. Unfortunately, I still see things like: "Cleanup would be ~84 separate removals. I didn't click through the UI to confirm." Uh, a) you're not clicking through ANYTHING opus (there is no web browser loaded), you were looking at some code. But you should read the code a little further and if you did you'd see what was going on. Fable gets lazy, but never like this. I've seen Opus do it twice in one thread. Not a huge deal, I nudged it back, and I assume it would eventually dig in when it came time to fixing it and would've explored further. "Where to remove duplicates. This is the one-way door. We can merge the overlapping rows in the worker, per image, or collapse to one reading per character per day in Rails." The "worker vs. rails" is a core architectural decision/rule that has been already ratified as law throughout the whole app and existing code - "we can merge in the worker" is RIGHT OUT and has never been an option. Again, something I've never seen Fable do; it knows that Option (A) - Worker - is completely wrong. Is this fatal? No, the worker could do the de-dupe, but it fails our design constraints and would introduce a whole bunch of fragility other places. I didn't let it get far enough to actually make a real plan; it's possible future iterations of the plan or a review check would've caught it. Again, a thing Fable does not do out of the gate, though. So I would say - I really like 5.5. It's way faster and cheaper than Fable. I don't trust it well enough (yet) to have it not break some major architectural rules in the app - but in full disclosure: I also never let it go far enough to see if it would right the ship. The risk I see - and I saw it with 5.0 all over the place - was it would make narrow decisions that were fine in the micro but it never zoomed out far enough to know that what it was doing was actually wrong. My concern (fear) is that continual training on "just run /goal and it will eventually beat the problem into submission" is fine if what you're doing is only that small problem. Trying to do that across large codebases will eventually result in working code - clicking the button does the thing you wanted it to do - but at the cost of broader cohesion and architectural stability and worse long-term maintenance. Fable rarely hits this problem; I hope both continue to improve in this area. That said, I am finding 5.5 is a great cross-check for Fable and my Opus allocation seems nearly infinite, so I am using it for all sorts of things where I'm not super worried about it getting something fundamentally wrong. And, Opus is so much faster that quantity is a quality all its own, and I am reaching for it when I'd usually use Fable (or even Astra). Absolutely great trajectory from Anthropic, though.

america.gov neat

Raise your hand if you AREN'T building an agent loop management system

I really need some Fable Stans to say "Opus 5.5 for long-range planning and architecture is as good as or better than Fable 5.1". That's what I need lol

Yes, yes Claude. That is ... very clear. "You owe one word: "released". Then §L runs: merge origin/main, linearize with the packet subject, force-with-lease, exact-head run, ready, auto-merge, land-verify by content, close hvu7.20, archive, reap."

"Astra hit AGI" if, like, you have infinite time and money and don't care that it will create a million line code monster just to drop two columns in a database table. Did it drop the columns? Yes. Yes it did. Did my test suite just 10x runtime and I now have 300,000 lines of code that will exist forever, yet never be called again? Yes. So, like, yes, if you do not care at all about the quality, performance, cost, wall-clock resolution etc. of the solution and just "did it complete it?" then, sure, AGI is here. But nobody would call this a good job or otherwise want the code this created.

I now burn out my entire weekly Max 20 in about 30-60 minutes of Fable. How nice.

The future is multi-modal. Anyone using GPT, Grok, and Claude extensively will see clearly where the weaknesses of one model are countered by the strengths in another. I always have Fable plan, GPT execute, and Grok critique. The architecture and code is far, far better (cheaper, too!) than using a single model family alone.

GPT strikes again! I need to learn my lesson and either a) never let it plan anything or b) really read the plan to figure this out before it spends 6 hours doing nonsense. I aint gonna do (b) so I will continue to use Fable to plan, GPT to do everything else

I've always found GPT much better at coding than even Fable, but where it lost me was on planning something. The only model that would go off the rails and build a rocket ship when I needed a skateboard was GPT. Fable was waaaay better at straddling the line between "this is too high level and handwavy" and "just enough detail to get it right without getting lost in the weeds". I found then GPT 5.6 Sol was always the best choice to implement Fable's plans. GPT would find all the holes and make Fable stop being lazy and go figure out what it should be. I'm sure nobody has had enough time, but wondering if anyone has seen anything with Astra - does it continue the "overcomplicated, too pedantic/cover-all-edge-cases-or-else, overbuild at all costs" attitude of Sol, or has it started to achieve the right balance?

Holy shit it’s AI Christmas