swyx

35 ms·

751.
by swyx 1y ago
> pen and paper is literally ten times more durable than even good digital storage. citation needed lol. data replication >>>> paper's single point of failure.
752.
by swyx 1y ago
i think usually deepseek posts a paper after a model release about a day later. no idea why they cant just wait a bit to coordinate stuff. bit messy in the news cycle.
753.
by swyx 1y ago
you can block grifters. easy. nothing you can do about absence of and harrassment of high signal AI people. which eugene is. that is bluesky's unique problem so far.
754.
by swyx 1y ago
we'd keep the curiosities around, like so much Ea Nasir Sells Shit Copper. we have room for like 5-10 of those per century. not like 8 billion. much of life is mundane.
755.
by swyx 1y ago
bluesky uniquely hates AI. very toxic around it. https://eugeneyan.com/writing/anti/ its sad that the informal social media bubbles we all have formally forked[0] into Red Twitter and Blue Twitter. just a direct c
756.
by swyx 1y ago
idk man how can URLs last forever if it costs money to keep a domain name alive? i also wonder if url death could be a good thing. humanity makes special effort to keep around the good stuff. the rest goes into the garbage collection of hi
757.
by swyx 1y ago
i get what youre saying but there are impt differences: 1) search is much cheaper to serve than chat 2) chat is a less frequent usecase 3) google paid for entry into a locked ecosystem (apple. also to appear like they are not a monopoly). t
758.
by swyx 1y ago
btw that podcast was recorded in jeff's office :)
759.
by swyx 1y ago
nice one simon - i'm guessing this is mildly related to your observation that everyone is converging on the same set of tools? https://x.com/simonw/status/1927378768873550310
760.
by swyx 1y ago
oooh this is >$1k mo talk
761.
by swyx 1y ago
oh i didnt actually mean for you to take actions, taglines are hard, startups need to claim ambitious comparisons, all analogies fail etc etc, love you dang its just we are currently confronted by a deluge of startups of "cursor for x&
762.
by swyx 1y ago
just 2c on the tagline but there was a time when this would be called "Copilot for Sales"... mixed feelings about what a better shorthand tagline would be but "Cursor for X" when there is no IDE, you're not forking
763.
by swyx 1y ago
i found it oddly reassuring/decontextualizing to search/replace Claude with "your outie" + its nice to read in a markdowny format https://gist.github.com/swyxio/f207f99cf9e3de006440054563f6c...
764.
by swyx 1y ago
> That last point indicates an interesting capability that diffusion models have: you get a kind of built-in quality knob. If you want fast inference at the cost of quality, you can just run the model for less time and end up with more n
765.
SWE Agents Too Cheap to Meter(latent.space)
3 points by swyx 1y ago|0 comments
766.
by swyx 1y ago
1) bad people can give useful advice. generally life is richer when you can separate art from artist. 2) 1 person bad does not make all people bad
767.
by swyx 1y ago
i understand the criticism but i also wonder at how to improve the status quo. Lethain is regarded as an authority because he has such a wide experience and writes so much and so thoughtfully. The truth is we just barely and rarely get such
768.
by swyx 1y ago
can we use this thread to share other useful HN tooling? mine is here: https://github.com/swyxio/hnx - displays the latest links from Y Combinator's Hacker News - Offers a link for submitting the url of the curren
769.
by swyx 1y ago
i think the key is always having the ability to telescope - think coding agents enable you to stay high level, but you always need the ability to go down and fix/understand code when needed.
770.
by swyx 1y ago
this is known as model routing in the lingo and yes theres both startups and biglabs working on it
771.
by swyx 1y ago
analysis as the resident summaries guy: - sonnet has better summary formatting "(72.5% for Opus)" vs "Claude Opus 4 achieves "72.5%" on SWE-bench". especially Uncommon Perspectives section - sonnet is a lot mor
772.
by swyx 1y ago
beyond 4 thats no longer true - marketing took over from the research
773.
by swyx 1y ago
good luck (genuine)
774.
by swyx 1y ago
thank you for the correction!
775.
by swyx 1y ago
livestream here: https://youtu.be/EvtPBaaykdo my highlights: 1. Coding ability: "Claude Opus 4 is our most powerful model yet and the best coding model in the world, leading on SWE-bench (72.5%) and Terminal-bench (43.
776.
by swyx 1y ago
i think the original intent of responses api was also to unify the realtime experiences into responses - is that accurate?
777.
by swyx 1y ago
(should add TLdraw which is specifically built for extensibility) no, it is overrated. there are many many canvas workflow builders like this one. they're all beautiful, intuitive and easy to use. they all also dont get traction. go ah
778.
by swyx 1y ago
people change, and you make mistakes. forgive and forget when the evidence changes.
779.
by swyx 1y ago
right now apples and oranges literally only because 1) unlimited unmetered use and 2) not in browser so async and parallel. like that stuff just trumps actual model and agent harness differences because it removes all barriers from thought
780.
by swyx 1y ago
this embeds words with word2vec, which is like 10 years old. at least use BERT or sentencetransformers :)

More ›