Kyle Hughes

😵‍💫 It’s launch day for Super Headache! This iPhone app is the fastest way to record headaches to Apple Health. There's just one screen and one big button. You can start recording now and come back when it's over. And it's completely free! apps.apple.com/us/app/id15640…

@steipete That would be Great-Depression-levels of unemployment, which probably means a significant increase in death by poverty and the end of class mobility. Why would that scorn not be deserved?

I wonder how long it’s gonna take till we get hate from normies as AI will replace 25% of the workforce in a few years.

How much water does Mastodon use every time you upload a video and it gets replicated into 10,000 different S3 buckets?

@jakehalloran1 @_xjdr Agree. It’s a yellow flag to me that the labs are now angling toward product-market fit, at least in part.

@_xjdr > the frontier is REALLY leaning in to vibe coding This worries me not from the perspective that code isn’t the best current use case for models but because it’s clearly the companies don’t have great monetization ideas other than vibe coding and there’s only so much there

the last week of launches has highlighted a few things for me: - Progress in LLMs has been amazing but incremental capability gains are clearly closer to log than linear while the corresponding cost for those gains is closer to exponential than linear. scale still works but the costs are scaling too. The next 18 months will be particularly interesting on this front - as we start to pivot to scaling RL flops, no one has really cracked efficiently scaling flops for inference time compute yet (still the most cost per flop efficient way to improve models). Google is closest from what i can see followed by OpenAI. Open source is far behind here. - I didn't look at benchmarks once when Opus dropped. I officially no longer care about current benchmarks. - in light of the death of (current) benchmarks, we have a dire eval problem. we can turn that frown upside down tho. we need much better evals for the big labs to hill climb [sic: goodhart]. the labs optimize for the evals and the evals optimize the labs right back. we should use this power - coding is the most powerful and important use case (other than automating math) and the labs seem to be focusing on unsupervised / async workers. the frontier is REALLY leaning in to vibe coding. that seems unwise (to me at least) but we will see how that plays out. maybe its just more bitter pilled than i am. - open source models are wildly underparameterized. o1 -> o3 feels like the same playbook with more flops (both in pretraining and RL) over more params. Benchmarks dont capture it but its very clear to see (and more importantly feel). - current heavy handed RLVR has a very steep tradeoff (see sonnet 4 / maverick) . we need better / different approaches. - no one has a good converged model (yet). for the foreseeable future we will see 'tool use / agentic' models, fast and cheap long ctx models, and 'general' models. This might be optimal as long as the 'general' models become more creative and interesting. i dont mind a few different models if they are actually different / great / interesting. - Open Source is so far behind on long context, vision, RL and post training polish. this primarily comes down to data quality and flops per param and is in no small part because DeepSeek hasn't _really_ invested here. Noam arch + MLA for global attention (i haven't spent enough time with NSA) or something equally ambitious should be the baseline and 1M ctx should be the minimum standard on the way to infinite context - we need (much) better open source RL envs. TRL and veRL are not going to cut it if OSS wants to actually chase the frontier. we also need better examples of how to properly do RL as scale. its very complex and unwieldy beast. im hopeful this is possible but this is the area if have the most concern - THERE IS SO MUCH LOW HANGING FRUIT it makes me feel like i am taking crazy pills we arent making more meaningful progress

@kimmonismus That this is how everyone in this space talks – breathlessly racing to claim that something is over or dead – assures me that UBI or wealth redistribution will never come out of the accelerationist community. It exists entirely on vindictive energy.

That would be the coffin nail for software engineers.

@mynameiscolton For some reason I have followed you on Twitter for most of my life. Never remembered why, never regretted it. Good luck, man.

hey twitter i’ve been in the hospital for 12 days… making slow progress back, but now am at the point of needing help to recover while i recover gofund.me/6c6e4ff1

Ya-ha-ha! You found me! @kyle_hughes" target="_blank" rel="nofollow noopener">threads.net/@kyle_hughes

@Leon_ADHS I assure you that if this ever becomes feasible I will add it to the app.

@Leon_ADHS Thanks for reaching out. Those are common requests I get. Unfortunately, a big focus of the app for me is not collecting any data: I rely completely on Apple to save and show the data. Unfortunately Apple doesn’t give me a way to show you any additional information ☹️

@merebemps your peloton workouts are inspirational

@alexisgallagher The Mammoth beta from @JPEGuin and the Ivory beta from @tapbot have been very well received. I haven’t tried either but they’re where most of the excitement is right now.

@alexisgallagher It’s a rich ecosystem. I really like Metatext: it has a “stock” UI feet, and great timeline-position memory and pagination. Unfortunately the developer recently health issues and stopped development. It won’t get post editing. Eventually I’ll switch to Ivory from Tapbots.

For the last month-plus I have moved all of my social media activity to Mastodon. It’s going great, I’m having more fun than I have in recent years on Twitter. I can’t tell you where to find me because of the draconian censorship here. Check my bio. I don’t plan to return here.

@alexisgallagher Toot! is one of the most popular, it’s very cute and polished. It has a wild (optional) threading UI. I would use it if not for Metatext. I really like the first-party web app on desktop. It’s sleek and fast and very customizable if you turn on the advanced interface.

@mattmedeiros @Pressable @kinsta thank you, sir 🫡

@KyleHughes Frustrations due to that price point point. Look to @Pressable and @kinsta if they don’t mind spending a bit more. They will even do migrations.

@mattmedeiros asking for a friend: what’s the best low-maintenance WordPress hosting solution? They’re on shared hosting on Dreamhost and very unsatisfied.

oof #MyTwitterAnniversary

@blackqueeriroh I agree with all of the complaints but most are multiple orders of magnitude less interesting and impactful than other things happening the space. Talking about bugs in old 3rd-party Mastodon apps on Twitter reads as whataboutism on behalf of Musk.

Some of y'all tweet like Mastodon was the billionaire that bought and skewered the community we built for 16 years.

@aaronmfisher I don’t know, I wash all of mine cold 😅 But I think most of their blends are not heavy on cotton so I wouldn’t expect much shrinkage. They hold up well in the dryer.

@aaronmfisher I don’t have a lot of their stuff because 💰 but the things I do have are all favorites. Definitely worth it. I packed both shirts for this trip!