great thread go read it
but if tldr models are improving 2x faster than 6 months ago
1. Benchmarks provide one piece of evidence for acceleration. Since Oct 2025, AECI scores for Anthropic’s models have grown at 1.9x [95%: 1.5–2.8] the earlier rate. Analysis choices affect this estimate, and it remains unclear how benchmark scores map to real-world capabilities.
15
but it does write like old opus, which smells
some thoughts on fable 5.2 vs opus 5.5:
the biggest difference is in how much attention/steering it takes to get things done. they talk similar, but fable is more pragmatic and actually steers subagents to finish stuff.
biggest pleasant surprise has been it detecting when the loops aren't adding value and dismissing things subagents find.
ultracode does a decent job of conserving fable limits - too early to tell but ultracode with sonnets might be the combo that makes the quotas last all week
14
upgrading auth for guv.sh
choice was between handrolling, self-hosting a lib, and using a managed solution
we're a small team so everything is about optimising attention
decided to go with the managed solution because we can't afford the distraction of patching all the time now that CVEs drop like hotcakes
LOC not that different, but attention burden would have been
everything goes in cycles, this is probably the same decision i would have made pre-AI because we didn't want to maintain more code
1
138
who's else has already used theirs 😂😭
We've given all Pro, Max, Ultra, and Team users one free full reset, across all models and limits. Use it anytime before December 10 at 2:00PM PT.
We recently fixed a few small issues that used more limits than intended, and the Fable resets didn't do much for those of you who'd already hit your weekly limits for all models. While these issues didn't affect everyone, we want to make up for the trouble.
You'll find it in Settings > Usage, or at clau.de/reset.
1
2
45
some thoughts on fable 5.2 vs opus 5.5:
the biggest difference is in how much attention/steering it takes to get things done. they talk similar, but fable is more pragmatic and actually steers subagents to finish stuff.
biggest pleasant surprise has been it detecting when the loops aren't adding value and dismissing things subagents find.
ultracode does a decent job of conserving fable limits - too early to tell but ultracode with sonnets might be the combo that makes the quotas last all week
my snowboarding 📈
The new @united nonstop from San Francisco to Sapporo, Japan, starting December 11, is going to be legendary for snow lovers.
It takes you to Hokkaido’s ski resorts, some of the snowiest on Earth, averaging 500–800+ inches of snow a year.
Cold Siberian air picks up moisture over the warmer Sea of Japan and dumps it on the mountains, creating a relentless powder machine.
24
the state of agents doing checkout is still awful. after an hour of back and forth, comparing different sites, finalising the bom, duet says it can't check the box and submit the order.
39
Adi retweeted
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
"treat any unnecessary context switch as a bug and fix it"
sigh
everyone i speak to is doing way more stuff, but feeling way busier and more fragmented. ai has massively lowered the activation energy of work: starting is cheap, finishing costs attention, so we do more in parallel and get less done.
the defaults of current tools/practices are akin to level 2-3 of self-driving cars, the messy middle between fully driving (flow state) and it driving itself so you get to actually focus on something else.
the game has always been spending attention efficiently. climb the autonomy ladder by handing off outcomes, not tasks:
1. define the highest-value repeatable workflows in your life
2. give agents the tools to work end-to-end
3. have agents organise and track work
4. reduce what reaches you to the decisions that actually need you
5. finish or kill tasks - an open loop costs even when you're not touching it
6. treat any unnecessary context switch as a bug and fix it
examples:
- i hate scrolling so i have several grokbot routines looking for interesting stuff. it started off having pretty bad taste so i scrolled for a while, tuned it by liking/bookmarking posts, and now it's decent
- i use omnesis to collect and index all my digital life on a mac mini. it took a week to set up, now i get daily summaries of what i missed across all my comms. no more jumping between 5 email accounts and chat apps to remember to reply to people
- i build guv.sh. users send us feedback on several surfaces: DMs, email, discord server, IRL. this gets automatically triaged into tickets and votes. if enough people + us have been talking about it, it gets picked up and worked on automatically. sometimes i wake up to a new slack channel with videos of an agent QA-ing a feature I was discussing last night. took weeks of chipping away and broke a lot along the way - now it's finally paying dividends
progress isn't linear, often it feels like extra work when there's already so much to do, but with a bit of consistent tinkering you can get things to level 4. it's the only way to get your attention back.
39
everyone i speak to is doing way more stuff, but feeling way busier and more fragmented. ai has massively lowered the activation energy of work: starting is cheap, finishing costs attention, so we do more in parallel and get less done.
the defaults of current tools/practices are akin to level 2-3 of self-driving cars, the messy middle between fully driving (flow state) and it driving itself so you get to actually focus on something else.
the game has always been spending attention efficiently. climb the autonomy ladder by handing off outcomes, not tasks:
1. define the highest-value repeatable workflows in your life
2. give agents the tools to work end-to-end
3. have agents organise and track work
4. reduce what reaches you to the decisions that actually need you
5. finish or kill tasks - an open loop costs even when you're not touching it
6. treat any unnecessary context switch as a bug and fix it
examples:
- i hate scrolling so i have several grokbot routines looking for interesting stuff. it started off having pretty bad taste so i scrolled for a while, tuned it by liking/bookmarking posts, and now it's decent
- i use omnesis to collect and index all my digital life on a mac mini. it took a week to set up, now i get daily summaries of what i missed across all my comms. no more jumping between 5 email accounts and chat apps to remember to reply to people
- i build guv.sh. users send us feedback on several surfaces: DMs, email, discord server, IRL. this gets automatically triaged into tickets and votes. if enough people + us have been talking about it, it gets picked up and worked on automatically. sometimes i wake up to a new slack channel with videos of an agent QA-ing a feature I was discussing last night. took weeks of chipping away and broke a lot along the way - now it's finally paying dividends
progress isn't linear, often it feels like extra work when there's already so much to do, but with a bit of consistent tinkering you can get things to level 4. it's the only way to get your attention back.
nhtsa.gov/sites/nhtsa.gov/fi…
levels of self driving, i find it a useful framework for automation in general
1
41
