sohail retweeted
It seems to me that democracy has on the whole worked very well in America over the last decade, and perhaps better than in any other major country in the West. National outcomes have repeatedly been responsive to underlying voter preferences on meaningful questions, and significant changes in policy have ensued. I’ve been reading Democracy in America this past week, and it’s striking how much it still rhymes nearly two centuries later.
sohail retweeted
I feel ever more confident this is an election with two waves: an anti-Trump wave and an anti-incumbent wave.
In areas with blue incumbents, they can cancel each other out. In areas with red incumbents, they build on each other.
everyone worried about OAI/ANT revenue are missing the fact that API revenue will be meaningless if they do get AGI/RSI.
sohail retweeted
Ask Claude or ChatGPT to give you a single random word. The answer is always "Lantern." Why?
if we have a pro-ai 2028 election, you'll know misalignment was never solved
teapaw.org/hilbertspaess/status/2…
personal ai assistants (a la instinct) are difficult to use/recommend because who knows what happens to your data.
connect my email and discuss personal matters and send to data centers in plaintext? no thanks.
sohail retweeted
Lack of transparency. Forcing communities to sign NDAs. Not knowing which company is building the datacenter. Not knowing if there’s a “gotcha” in whatever agreement you’re signing. Worrying that datacenter companies won’t follow through.
sohail retweeted
I built something I’ve wanted for a while: a Techmeme/Memeorandum for China.
China Desk pulls from Chinese sources to surface what’s getting attention across politics, culture, AI and tech.
Link in thread 👇🏼
why does opus feel bad?
- the writing. the writing is horrible. ask it to write you a doc and it just writes the most semantically dense garbage that is basically unreadable
- agent usage. my workflow is to take feature work, break it up, have it research using subagents and ask me questions before using subagents for implementation.
if will often just stop using subagents after asking questions or finishing its first series and i'll have to stop it and prompt it to do so. it'll just blow up the context otherwise
- compaction. codex feels so far ahead here. i'm constantly blowing past whatever claude code thinks is the max context should be and cc doesn't try to compact at all.
feels like some of these issues are model related while some can easily be fixed with a better harness.
sohail retweeted
i've got a handle on all three now. which one you reach for depends on the work in front of you.
sol. digs in and doesn't come up for air.
a weird bug, a problem you have to grind at. hand it over and it goes all the way down, further than either of the other two. it holds up on a full plan too, just a step behind fable. ask it for a new angle and it has nothing.
→ give it: one point, all the way down
opus 5. most ideas, least depth.
ask it for angles and it hands you a pile, second only to fable. push it two levels down and it drops off, well behind the other two.
→ give it: openings and groundwork
fable 5. hand it off and stop watching.
a step behind sol on depth, best of the three on breadth and on coming up with angles. nothing about it is scary good, but you can't poke a hole in it either.
→ give it: the job you can't describe the shape of
try running all three together.
no reason to run 𝗳𝗮𝗯𝗹𝗲 𝟱 anymore.
it's 5.6 sol and opus 5 tag-teaming now, and that combo covers way more than i expected. opus 5 is the more complete one once you count its frontend work, but for some reason sol just feels more dependable.
the chart ranks them like this, but a few things are worth noticing:
→ the top four don't separate. 74±4, 73±3, 70±4, 70±3. the bars all overlap, so 𝘁𝗵𝗲 𝗿𝗮𝗻𝗸𝗶𝗻𝗴 𝗶𝘀𝗻'𝘁 𝗮 𝗿𝗲𝘀𝘂𝗹𝘁.
→ sol runs 61 steps and 60k output tokens. opus 5 needs 99 and 118k. same score, 𝟯𝟴 𝗳𝗲𝘄𝗲𝗿 𝘀𝘁𝗲𝗽𝘀, 𝗵𝗮𝗹𝗳 𝘁𝗵𝗲 𝘁𝗼𝗸𝗲𝗻𝘀. among the seven models above 67%, sol takes the fewest steps.
→ fable 5 and opus 5 do still overlap on error bars, but it's $21.63 a task against $11.84. call them tied and it's still 𝟭.𝟴𝟯𝘅 𝘁𝗵𝗲 𝗺𝗼𝗻𝗲𝘆.
→ fable 5 puts out 119k tokens over 88 steps. opus 5 118k over 99. same work, fewer steps even, and fable 5 still bills 1.83x. 𝘁𝗵𝗮𝘁'𝘀 𝘁𝗵𝗲 𝗽𝗿𝗶𝗰𝗲 𝘀𝗵𝗲𝗲𝘁.
would be great if we could get keyboard shortcuts to quickly change the model
@thsottiaux @Dimillian @dkundel
sohail retweeted
In spite of our many problems and challenges, I will forever remain bullish on a country that produces people like Victor Glover.
Victor Glover (pilot of Artemis II) was already a spacecraft pilot for SpaceX Crew-1. He is a naval captain who has had his naval aviator wings for 25 years. He has three (3) masters degrees, including in flight test engineering and systems engineering. He's been a test pilot since 2007. He has thousands of flight hours.
sohail retweeted
as a very, very proud american citizen i stand with dario amodei and appreciate his willingness to defend the civil liberties of other american citizens 🇺🇸
OAI's models are def better at the 'meta AI' aspect of the race atm. it'll be a competitive advantage imo
gpts are trained on python and are much better at designing ML experiments/writing prompts.
meanwhile, anthropic is all in on terminal use and gemini is general knowledge.
