Making games and robots and sometimes other things. (Formerly: 1X, Meta)
Bay Area
Joined January 2009
- Tweets26.9K
- Following372
- Followers1.6K
- Likes38.5K
Pinned Tweet
I wrote my first substack piece about AI's economic impact, and some reasons why models seem to underperform in the real world.
Andre Infante retweeted
my vocal synth models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings. still working on improving naturalness by tuning it against recordings of real singers, but I'm already impressed at how cool it sounds!
Andre Infante retweeted
METR has released a very interesting report suggesting a recent acceleration in AI capabilities this year at Anthropic due to RSI.
But they can't confidently reject an alternative explanation: the step change you get with the larger pre-train of Mythos.
I keep telling myself I'm gonna stop getting bogged down in these fights, but... "would be incels" is a ludicrously cruel and uncurious way to refer to children who dared to enjoy a book you didn't care for.
I have to remind myself that this sort of piece is functionally just a kind of ragebait and makes money by upsetting people with ludicrously bad takes so the piece is shared, clicked, and discussed.
The new Gemini being frontier competitive in raw intelligence seems like pretty direct evidence against a 2x uplift from internal models. Would require the Gemini team to be twice as good as OpenAI and Anthropic, which I see little evidence of.
If you try even a little bit to gender swap the complaint, it becomes super clear that it's just a very inappropriate way to talk about the reading preferences of children.
Andre Infante retweeted
Should you donate $10,000 to vaccinate 2000 infants, or give TB medicine to 1000 adults? Wrong question - you can't judge lives like feathers on a scale. Priorities should be decided on democratically. You should buy one-fiftieth of a Patriot missile for war with Iran.
The way normal people think about ethics is honestly pretty bleak, to me. It seems to almost purely vibes and social posturing. "Oh you're trying to think rigorously about how to do the most good? That's a big yikes from me, dawg. Read the room, buddy."
Readers added context they thought people might want to know
The excerpt is from Dario Amodei's 2010 GiveWell guest post as a Princeton physics grad student; he donated the $10k to VillageReach (child vaccines) and described the critical analysis as emotionally taxing.
blog.givewell.org/2010/06/03/my-…
nytimes.com/2026/11/05/tec…
Maybe you don't agree with him. Children give up more remaining lifespan when they die, you could build an argument about that. But if you think the idea of entertaining the question is somehow offensive, it's because you're stupid and can't think.
I think daybreak and glasswing might be kind of bad. The motivation there is to minimize misuse of frontier models. But we don't actually care about that, we care about the balance of offense to defense.
Supervised whitelist programs with manual and exclusive approvals allow them to be very heavily loaded towards defense, but at the cost of them being *used* vastly less. We might well be better off tolerating more misuse in exchange for vastly more use overall.
Even ignoring the slightly longer term, doing this unilaterally would be remarkably suicidal given just the cyber security stuff that is already happening right now. Zero imagination or insight required.
I'm very happy that there is now a check on Republican power, but I have very little faith in the ability of Democrats to competently govern wrt to AI rather than trying to score populism points and punish disfavored groups. I see worryingly few signs of seriousness.
This really does not comport with my experience of Astra, at least in the domains I'm using it in.
I've had a couple of experiences where Astra wasted hours or days piddling around on a problem and then I gave it to Opus 5.5, which just ~immediately figured it out.
With better uncertainty estimates and a joint distribution over all models’ ECIs, we can now compare models directly.
For example, we can now say Astra has a 72% probability of being more capable than Opus 5.5, but a 99% probability of being more capable than Fable 5.1.
ALT Dot-and-whisker chart of the ECI gap between GPT-6 Astra and each of the five next-highest models under the new method, with 90% intervals. The probability that GPT-6 Astra is more capable is 54% against GPT-6.1 Sol, 72% against Claude Opus 5.5, 99% against Claude Fable 5.1 and Claude Sonnet 5.5, and 100% against Gemini 4 Argon.