Cloudbring (π,π) retweeted
The METR graph, updated:
Readers added context they thought people might want to know
This chart is not the official METR graph. It was created by Benjamin Todd via extrapolation from Anthropic's internal AECI scores, as METR is no longer updating its public time horizon measurements.
x.com/ben_j_todd/staβ¦
benjamintodd.substack.com/p/ai-trend-extβ¦
metr.org/time-horizons/
metr.org/horizon-chart-β¦
metr.org/blog/2026-1-29β¦
Remember how concerned we were with deep fakes and how people wouldn't be able to tell they were looking at one?
28
Me: make it 10x faster
Claude: I'm sorry I could only make it 6x faster so far. I'll continue working on it.
20
A lot of Big if True energy on this one.
Today weβre introducing Zyphra Hierarchical Attention (ZHA), a path to true subquadratic attention.
ZHA searches a learned tree of compressed summaries, with projected speedups of 35Γ over dense attention and 5Γ over MSA at 10M tokens, increasing to ~400Γ and ~55Γ at 1B. π§΅
24
Today weβre introducing Zyphra Hierarchical Attention (ZHA), a path to true subquadratic attention.
ZHA searches a learned tree of compressed summaries, with projected speedups of 35Γ over dense attention and 5Γ over MSA at 10M tokens, increasing to ~400Γ and ~55Γ at 1B. π§΅
You know AI is truely at human level not when it can do one recognizable style of output at human level but, when it can do multiple styles at human level and not one AI aesthetic. Then it will truely be super human level when you can tell from the aesthetics itself that no human could have pulled that off.
11
Just use ULTRATHINK when you want Claude to pay attention and medium the rest of the time.
If you need to raise it, you're probably prompting badly which in the end of 2026 means you're not giving a solid goal.
how often do you change reasoning levels mid thread?
12%every other thread
16%like 3 times a day
33%maybeee 1 per day?
39%i cant remember
1,617 votes β’ Final results 20
Claude, if you're so smart, instead of math problems or curing cancer, how about you stop the 5 parallel rust compilations from crashing my mac mini.
Just built a homelab around T3 code's ability to have remote workers on a server. Looking forward to moving concurrency up from 3-5 on my mac mini to 8-12 on my server.
My rust port is now limited by the number of cores I have on my dev machine. Building a 16c/32t 128GB linux machine just for my agents to build and using T3 Code to run it this weekend.
Claude, if you're so smart, instead of math problems or curing cancer, how about you stop the 5 parallel rust compilations from crashing my mac mini.
23
Next we'll hear announcements of Philosopher's Stone and finding the Fountain of Youth
Replying to @MarathonFusion
In 2025, we published a bold proposal to use fusion reactors for gold production.
And now weβve done it: Mercury transmuted into stable gold, using fusion neutrons, at 10x the record set by CERN in 2025.
Weβve reduced alchemy to a series of difficult but tractable engineering problems.
8
Cloudbring (π,π) retweeted
People keep using short timelines as an excuse to do short-term research.
But if legions of AIs will be helping us soon, we should pursue research plans that intuitively feel like theyβll take many decades.
I say βan excuseβ because the latter seems bottlenecked on courage.
We're only a year into this camera and the already we are building impossible visual languages.
2
29
There is no longer any moat in games. The relationship the game makers have with fans is the only thing left. Everything else can be modded.
1
10

















