B Webb retweeted
Carlos comes alive😤
In the last 4 games of the 2nd set from 5-3 down, Carlos wins 50% (8/16) of points from defense 🏃💨
(Tour avg - 33%)
B Webb retweeted
The meeting between Pope Leo and Dylan was a brief one, consisting of greetings and a few lighthearted exchanges, according to Vatican News.
Pope Leo presented Bob Dylan with a rosary, while Dylan reciprocated with a book: Chronicles: Volume One.
Point of the match...and possibly the week 🔥
@carlosalcaraz simply on fire in the final thus far
#ErsteBankOpen
A new session starting with
@mbw61567742 articles
is live here. claude.ai/share/2469509a-99e…
A 3-way scenario is currently P(LL)=0.17%, P(market zoonosis)=92.5% P(non-market zoonosis)=7.2%.
Report here: drive.proton.me/urls/ZENMGEN…
LL side includes: MBW's papers, and substack article; @vbruttel on RE sites; Defuse proposal; Levin; Stoyan/Chiu; @ayjchan op-ed; @ydeigin on FCS;
@mattwridley on Laos claims; Daoyu Zhang @gdemaneuf et al. '22; @stevemassey, @stevenquay and Lv et al.; pre-adaptation papers
1
1
217
B Webb retweeted
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
On the use of Claude Sonnet/Opus to judge Covid origins. This was done in two ways in closed book form i.e., only based on the provided documents. In the first one, Claude gave a P(Lab Leak)<0.06% based on the Rootclaim debate alone. teapawt.net/tgof137/status/2103595…
Here's one more interesting AI judging result, courtesy of @bratwebb and Claude.
drive.google.com/file/d/1OFl…
This used a two pass approach where it first extracted all the arguments, then did a Bayesian analysis where it seemed to pick very reasonable numbers.
Adding Claude's complete rejection of Bruttel's RE site arguments in its entirety. teapawt.net/bratwebb/status/210501…



















