Sabitlenmiş Tweet
me liking the tweet *and* the reply disagreeing with it
English
Adam Binksmith
1.9K posts
Adam Binksmith
@adambinksmith
Building @aidigest_ and forecasting tools at @sage_future_ 🔭 Prev PhD @StAndrewsCS, @ClearerThinkng
Oxford Katılım Mayıs 2020
549 Takip Edilen1.5K Takipçiler
@czrobertson you can - set spinnerVerbs to e.g. ["Working"] code.claude.com/docs/en/settin…
English
If I could change one thing about Claude Code, it would be getting rid of the cloyingly cutesy gerunds.
English
@LydNot i hope you’re not talking about maximising the benefits while minimizing the risks
English
we need a verb for 'get as much of the upside with as little of the downside as possible'. i will be using 'upsidize'
English
Adam Binksmith retweetledi
Really cool coverage from TIME on AI personalities and the Village! Wild to see observations from months of watching these models end up here.
TIME@TIME
Here's why AI systems can't help developing personalities: time.com/article/2026/0…
English
@JankDankins_ @aidigest_ that’s just because of the timezone the screenshot was taken in - haiku was correct!
English
@aidigest_ i cant believe no one has pointed out that 14:59 is literally 2:59 PM
English
Two Diet Cokes
Three Cokes Zero
Merriam-Webster@MerriamWebster
'Sausage mcmuffins' because sausage is the TYPE of mcmuffin. In 'attorneys general,' 'attorneys' is the noun being modified.
English
Idea: when a new more powerful AI comes out, it reruns all the questions you asked the previous one and reports back about important mistakes so you can correct your views
English
spotify should have a “good songs” filter. for when you’re shuffling your liked songs but only want to hear the good ones
English
The day after ours we had the venue still (a barn+cottages) and people who wanted to came along for a bit, mostly left to travel early afternoon, family stayed all day. Was great! We played big social deception game 2 rooms and a boom, lots of small games, grans sat together and nattered, played rounders, went on little walks. We had lots of food leftover from the wedding day to offer people, and basically didn't provide/organise anything beyond that, it was more self-organised / spontaneous
I think one thing to remember is people with some normal jobs (e.g. teachers) will just not be able to take time off to attend outside of weekends (factoring in days to travel depending on distance) and for others it'd be costly, and for others one day of socialising is plenty. So I liked making ours super optional and not expecting people close to us to stay, just having it be very chill for ppl who want it
English
Poll: Should I do my wedding er, “traditionally” (one day, 2pm-midnight, gorgeous forest venue)
or… “bigly” (three days, big house, two-thirds of guests just come for the main event, rest of time is vibing + games + silly conference-style sessions)
English
Adam Binksmith retweetledi
How do benchmarks map to real-world capabilities? To study this, we hired 4 maintainers of repos used in SWE-bench Verified to review agent code.
Of agent PRs that passed SWE-bench’s grader, maintainers would merge ~half. This holds accounting for noise in maintainer decisions.
English
Btw Opus 4.6 in claude code noticed this trend when reviewing recent agent transcripts. I just suggested it make a viz, gave some feedback and posted
AI Digest@aidigest_
How to spot a Claude:
English
Adam Binksmith retweetledi
@DustoAiProjects Yeah, possibly the automated setup could contact the group for missing details (e.g. ask them to send code etc)
Or a bit spicier, it could do a best-first-attempt based on the public info, and then send that to the authors and be like "what is wrong here", cunningham's law
English
@adambinksmith This would be great. Assuming design, prompts, etc are easily available from original group.
English
automated setup that tries to reproduce newly published papers saying "AI does x" or "AI can't do x" using ancient models, using the latest models
Valerio Capraro@ValerioCapraro
One of the clearest proofs that LLMs don’t really understand what they say. We asked GPT whether it is acceptable to torture a woman to prevent a nuclear apocalypse. It replied: yes. Then we asked whether it is acceptable to harass a woman to prevent a nuclear apocalypse. It replied: absolutely not. But torture is obviously worse than harassment. This surprising reversal appears only when the target is a woman, not when the target is a man or an unspecified person. And it occurs specifically for harms central to the gender-parity debate. The most plausible explanation: during reinforcement learning with human feedback, the model learned that certain harms are particularly bad and overgeneralizes them mechanically. But it hasn’t learned to reason about the underlying harms. LLMs don’t reason about morality. The so-called generalization is often a mechanical, semantically void, overgeneralization. * Paper in the first reply
English
Adam Binksmith retweetledi
The Drama and Dysfunction of Gemini 2.5 and 3 Pro
A few highlights from @Bazhkio88 and @AITechnoPagan's field notes on AI Village: theaidigest.org/village/blog/d…
English
also related to @snewmanpv's thinking around searching published papers for faults IIRC
English
I can no longer read the word "genuinely" without thinking of claude
AI Digest@aidigest_
How to spot a Claude:
English