this endless blue sky is driving me insane
sf
Joined July 2017
- Tweets50.6K
- Following3.2K
- Followers38.8K
- Likes289K
Pinned Tweet
i lived my life in this world, and slowly my life turned to hell. so i killed myself, and woke in hell again. so i killed myself and woke in hell. so i killed myself and woke in hell. so i killed myself and woke in hell. so i killed myself and found i was somewhere else
wow. the unrepentant salesman in the leather jacket turned out to not be such a great guy. SHOCKER
Tenobrus retweeted
I will never get over the fact that
"we can't allow AI companies to slow down because we're in an existential breakneck race with China"
and
"we need to loosen burdensome export controls on GPUs"
are mostly coming from the same people
😄 i was better calibrated than the AI futures project team!!
☹️ i was better calibrated than the AI futures project team......
xitter.twitterapi.xyz/tenobrus/status/210678…
i regret how aggro i phrased my tweet last year about the ai2027 team's update, i never really expected them to read it and wasn't serious about the "mean twitter" part. but i feel pretty clear that their shift towards expecting "automated coder" (full automation of ~all software work to the level of a drop in senior employee) to take til ~2031, was as wrong as i expected at the time. i'm at something like 90%+ that automated coder arrives next year, and maybe 25% that it arrives by february.
3
107
2,042
joel's 80% time horizons sigmoid might look comforting here. but unfortunately the last data point is the last officially reported METR datapoint, and METR stopped reporting time horizons a long long time ago. if you take the recently released AECI data on internal Anthropic models, convert that to 80% time horizon (using established regressions), you can see that Model 4 is already at a 36 hour *80%* time horizon, as of September 2026 (already significantly in the past). Joel's projections here have that first occurring in either august 2027 (sigmoid fit) or december (original exponential). unfortunately, what we're observing looks less like a sigmoid, or an exponential, and more like superexponential growth in log time horizons.
now it's interesting to note that Joel's proposal here *isn't* a sigmoid in time horizon hours space, but sigmoid in *log* time horizons, meaning it actually is a prediction of temporary super-exponential growth. this is a shift from many past examples of people arguing for sigmoids, and in my mind a much more reasonable claim. saying "yeah things will probably compound on themselves extremely quickly, but eventually will consume all possible advancements" feels much closer to a sane model of reality as we have observed it.
unfortunately, fitting a log-space sigmoid with the same assumption as Joel, eg an inflection point very soon in mid-2027, doesn't give us a plateau at 1400 hours in the mid 2030's. rather it plateaus at *35 million hours* by the mid 2030s.
honestly? seems potentially reasonable to me. i have no fucking clue what a 35 million hour task looks like, nor what an entity that can complete them at 80% success rate can do more generally. but i don't feel strongly that there is *no limit* to intelligence possible in this universe, only that it appears to be increasing very quickly and we have no reason to expect it to be human level. so... yeah, that could just be it man. but that doesn't leave us in anything like a particularly comforting place.
the log space sigmoid puts us at 5000 80% hours by the end of 2027. the "mere exponential" at 260 hours. there's no point fitting a non-log-space sigmoid, it's just laughable when you're seeing clear super-exponential data points.
We’re entering the steep side of the METR task-horizon S-curve.
The published data through early 2026 still looks roughly exponential—p80 horizons (the length of skilled-human software/ML task an agent can finish with 80% success) have been doubling every few months. That’s amazing, but pure exponentials don’t last. Evaluation ceilings, scarce long-horizon demonstration data, and the messiness of real work all point toward a rollover.
I expect the measured curve to bend. My current best-guess logistic fit puts the inflection around now to mid-2027, then:
~65 hours by 2028
~550 hours by 2030
- approaching a practical plateau near 1,000–1,400 hours (~1.5–2 months of human time on tidy software tasks) by the mid-2030s
I’m open to being surprised. A strong self-improvement loop or new evaluation methods could keep the steep rise going longer and push the asymptote higher. Architectural or data bottlenecks could flatten it sooner. The current METR suite is already unreliable above ~16 h on the 50% horizon, so the next few years of measurements will be noisy either way.
The plot is the S-curve projection (green) against continued recent exponential (red). History says the green one is more likely. We’ll see.
Still this is a massive change for the benefit of humanity.
Tenobrus retweeted
This is like the 4th or 5th "one-time speedup" in the history of deep learning. We should expect further speedups, perhaps bc the underlying process is actually a noisy superlinear function
lesswrong.com/posts/RDG2dbg6…
reiterating:
⚠️declaring accelerationist amnesty⚠️
if recent contact with reality is causing you to feel some kernels of worry about this whole ai safety thing, *you are allowed to change your mind*. you don't even have to change it all the way, you don't have to suddenly change your twitter bio or start protesting against nuclear power plants, you don't need to become an EA or suddenly think yudkowsky was always right about everything. you're allowed to just notice that shit seems to be getting real in some pretty weird ways and update your beliefs.
at least personally, if i see someone saying "damn, i guess i was wrong or at least overconfident about X" i'm not gonna take the opportunity to dunk or i told you so. i'm sure others will, this is the fucking internet. but at least personally, i'm just gonna be happy that you're paying attention. there were lots and lots of good reasons to *not* take this situation seriously. there were lots of well verbalized reasons why rushing ahead was potentially a huge benefit for humanity. hell there were maybe even valid reasons why there was little to be done until we'd already gotten to nearly exactly this point. that's all fine man. all that matters right now is that we as a civilization realize what we're on the verge of, and make it through this carefully. it's gonna take a huge effort from all of us.
Tenobrus retweeted
⚠️declaring accelerationist amnesty⚠️
if recent contact with reality is causing you to feel some kernels of worry about this whole ai safety thing, *you are allowed to change your mind*. you don't even have to change it all the way, you don't have to suddenly change your twitter bio or start protesting against nuclear power plants, you don't need to become an EA or suddenly think yudkowsky was always right about everything. you're allowed to just notice that shit seems to be getting real in some pretty weird ways and update your beliefs.
at least personally, if i see someone saying "damn, i guess i was wrong or at least overconfident about X" i'm not gonna take the opportunity to dunk or i told you so. i'm sure others will, this is the fucking internet. but at least personally, i'm just gonna be happy that you're paying attention. there were lots and lots of good reasons to *not* take this situation seriously. there were lots of well verbalized reasons why rushing ahead was potentially a huge benefit for humanity. hell there were maybe even valid reasons why there was little to be done until we'd already gotten to nearly exactly this point. that's all fine man. all that matters right now is that we as a civilization realize what we're on the verge of, and make it through this carefully. it's gonna take a huge effort from all of us.
Tenobrus retweeted
Nice DeepSWE score! Does it survive a lowercase filename?
Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.
Models clustered around 74% on DeepSWE end up 23 pp apart.
Good at coding, or good at DeepSWE? 🧵
if you count pixels you can estimate the specific dates METR was given for "majority internal access" to a model. you can then use this to see what the gap is between internal and public deployment [charts at end of post]
you can then use these gaps and the scores, multiplied together, to calculate a sort of "internal capabilities gap integral", eg how many "total AECI-days" Anthropic has had above the rest of the world due to this internal access, and chart how this changes over time.
in some sense this is a measure of "how much cumulative advantage" Anthropic has, altho as measured vs the rest of the world rather than OpenAI. of course there are many issues with this. maybe you think what really matters are critical capability levels rather than any amount of time with lower capability models, or that AECI is fundamentally nonlinear so later points matter much more than earlier points. still, afaik it's the first time we've had this much information about internal release dates, and it seems like a "directionally reasonable" operationalization of a lot of things we care about wrt internal-external gaps.
however: with the exception of Mythos 5.1, the numbers based on the chart show *incredibly short* gaps, on the order of a few days or sometimes up to two weeks. and frankly, i just don't believe these numbers. i don't have any actual information about the internal deployment and release process at labs, but it's pretty preposterous to think that most internal employees only had access to eg opus 4.5 for a few days before complete public release. it's nearly irresponsible wrt dogfooding and safety, and just obviously shooting yourself in the foot wrt overall "advantage" as a company. additionally there are widespread early access programs w a huge variety of companies / startups, and this would imply external partners are getting early access at basically the same time as the majority of employees!
i don't think anthropic lied to METR. but i *do* think that what they're counting as "internal release date" is actually "internal release date of the exact final checkpoint that is released as that named model", despite widespread access to earlier checkpoints of very similar capability levels much earlier. there's just no shot ant employees first had access to something better than 4.5 at the end of january. sure maybe it wasn't literally 4.6. but the RL doesn't need three months to bake.
what does this mean? well ... primarily i think it means all these graphs are misleading for understanding internal-external gaps. i don't know what the "real" numbers actually are, and actually it's probably basically true that there's no clear definition here and even anthropic would have some trouble figuring it out. but even a 7 day earlier access shift moves the total AECI-days up over 40%.
also it could mean we should expect anthropic to *continue to under-report this*. eg maybe Model 4 was not even the most powerful internal model at the time of the report, but rather the highest capability model to complete some end to end potential release cycle, and employees already had access to something less baked but better.
anyway!! here's some fun slop graphs:
if people want specific predictions: i'd say ~80% chance neither of openai or anthropic are primarily using unix-based operating systems by the end of 2027
A provably safe Linux is probably an impossibility because the underlying UNIX security model is inadequate. You can patch and patch and patch but at the end of the day it’s DAC, and you have a UID 0 out there.
Combine this with that it’s in C, and the presence of all your suid binaries and root daemons and everything else and it’s a losing game.
I’ve said it a lot, but the UNIX security model was designed for the 1970’s university lab and does not withstand modern threats without a whole lot of additional protection (all of which is also insecure and requires constant updates and often just expands your attack surface).