Daniel Lee (@dylayed) profile picture

개발 이야기를 좋아합니다.

San Francisco, CA
Joined September 2015
Photo by Daniel Lee (@dylayed)Photo by Daniel Lee (@dylayed)Photo by Daniel Lee (@dylayed)Photo by Daniel Lee (@dylayed)Photo by Daniel Lee (@dylayed)Photo by Daniel Lee (@dylayed)Photo by Daniel Lee (@dylayed)Photo by Daniel Lee (@dylayed)Photo by Daniel Lee (@dylayed)Photo by Daniel Lee (@dylayed)
Daniel Lee retweeted
Check out my talk at Cloudflare Connect on celld. celld let’s you self-host workers, durable objects, queues, and workflows using only an object store bucket as a backend. We’re in the process of merging the ideas in celld into the open source workerd. More to come on the new workerd soon! cloudflare.tv/shows/cloudfla…
24
52
5
571
24,026
Daniel Lee retweeted
I was wrong about agent sandboxes. In my article I said pattern 1, where the agent runs on your infra and only calls a sandbox for tools, doesn't work at scale. I moved everything to pattern 2 and put the whole agent inside a VM. At the time that made sense but today I'd pick pattern 1. For your own personal agent, a coding agent, or a small project, isolating the whole agent in a VM is still a great default. I would keep doing it. At real multi-tenant scale it falls apart. There will be more agents than humans. Every human gets one computer. Giving every agent its own VM running 24/7 means billions of CPU-hungry machines, and we're already short on compute. Most of what an agent does doesn't need a machine at all. It thinks, calls APIs, reads pages, and waits. Nobody needs a 4 CPU VM sitting idle so an agent can run a simple JS script once an hour. What I'd build now: > the agent loop runs on shared infra that scales > it gets a sandbox only when it has to run untrusted code > sandboxes spin up in seconds and die right after > zero state on the machine, everything lives outside it VMs are a tool the agent calls, but they shouldn't be where it lives.
70
28
15
375
45,290
Oh shit we are arguing about monorepos again! Tech twitter is healing.
Monorepos were always an anti-pattern
1
9
776
Daniel Lee retweeted
If you're building with agents, or thinking about software factories, don't miss this new post from Matthew Goldberg about the new AWS Emulator: aboutamazon.com/news/aws/how… There are two very cool things here.
9
21
1
169
9,692
Gemini Decision 1 from @GoogleAI is now available on AI Gateway. Get structured decisions with probabilities from text, images, docs, video, and audio. Use it to classify, route, and score. 𝚐𝚘𝚘𝚐𝚕𝚎/𝚐𝚎𝚖𝚒𝚗𝚒-𝚍𝚎𝚌𝚒𝚜𝚒𝚘𝚗-𝟷 vercel.com/changelog/gemini-…
7
1
3
49
4,444
Daniel Lee retweeted
So why doesn’t glue work get you promoted, if it’s so crucial to shipping projects? Are companies stupid? Are they deliberately leaving value on the table? No, I don’t think so. Companies don’t reward glue work because they don’t want you prioritizing it.
17
14
5
416
38,800
DeepSeek's DSH harness is built on a simple idea: everything is a plugin. Including the agent loop itself. In this tutorial I break down cordis, the plugin system behind it: > the architecture of dsh > services, dependency injection, and events > building a real plugin live Thread 👇
17
31
2
417
16,365
My entire feed rn
🤖 Made with AI
my vocal synth models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings. still working on improving naturalness by tuning it against recordings of real singers, but I'm already impressed at how cool it sounds!
17
200
11
3,224
65,934
Daniel Lee retweeted
생각보다 보컬로이드가 내 취향이였나
Oh my god it can sing in Korean too
73
7
288
14,971
생각보다 보컬로이드가 내 취향이였나
Oh my god it can sing in Korean too
73
7
288
14,971
Oh my god it can sing in Korean too
my vocal synth models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings. still working on improving naturalness by tuning it against recordings of real singers, but I'm already impressed at how cool it sounds!
6
335
39
1,128
59,785
Daniel Lee retweeted
my vocal synth models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings. still working on improving naturalness by tuning it against recordings of real singers, but I'm already impressed at how cool it sounds!
I only make serious software
639
2,600
615
23,468
3,120,202
Stop forcing a 2,700-word prompt + 900ms frontier model to do a fine-tuned smaller model’s job. 💪 @AirbnbEng just fine-tuned @GoogleGemma 4 E2B on a single GPU with @vllm_project, to serve as a JEV-like classifier. The results are pretty solid: ⚡ 35ms median latency (85ms p99) 🎯 0.902 F1 (matches frontier models) 📉 ~1% of the cost at scale 🔗 Read the full deep dive on GLOWC + their @GoogleDeepMind Gemma 4 setup here: airbnb.tech/ai-ml/frontier-q…
17
52
2
432
23,941
Daniel Lee retweeted
The engineer who says no all the time is a real archetype among senior and staff engineers. Their role is to slow things down, to block the development of features that add complexity, and to ensure that as little code gets written as possible (since code is a liability).
33
50
15
1,025
88,914
Daniel Lee retweeted
check out my new article in the Anthropic eng blog. i talk about "dreaming" (offline compute) as a way to organize Claude's memory. it can correct mistakes, create memories, or build new representations. xitter.twitterapi.xyz/AnthropicAI/status/211…
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. anthropic.com/research/compo…
59
55
12
627
36,636
Daniel Lee retweeted
I unironically agree with @bcherny’s take that “coding is solved, but bugs are not.” The deeper meaning is that you can in theory code *anything* now. You can burn infinite tokens refining and optimizing. Including optimizations that no one will care about 💀 With Vercel you can build a different kind of recursive improvement loop. You capture traces from production, you then optimize and course-correct based on how your software is being used IRL. This is not a new idea. The best engineers I’ve ever worked with never jumped into coding (now prompting) right away. You instrument the system, collect data, understand distributions, user journeys, and then put fingers on keyboard. You can ask your agent today to look at your traces!
152
26
20
619
54,136
Vercel Tracing is now GA on all plans. Starting with Next.js 16.5, new projects have it on by default. Search and inspect spans from the dashboard or the CLI. ▲ ~/ vercel traces search --status error vercel.com/changelog/vercel-…
14
14
10
204
64,062
Daniel Lee retweeted
Just rolled out remote agents in Antigravity so now you can just fire and forget anywhere. It’s been the way we’ve been working at Google for some time so excited for you all to give it a go! More to come tomorrow 👀
85
48
28
941
111,173
Daniel Lee retweeted
my entire career, i have been warning people to not trust Linux for local privileges. i have always considered the VM boundary to be the minimum security for untrusted code, and limited the permissions of the VM (via things like IAM) to the smallest amount, assuming that any compromise of the software will compromise the entire system it may turn out that the security work being poured in vulnerability discovery is the first time that we get to a ~"provable safe" Linux, in which -- suddenly -- running *containers* is actually very safe (and cheaper than microVMs) because we *know for certain* that local privilege escalation is not possible this would be a big, surprising change and i would need to update how i think about building software
we've found, validated, and prepared patches for almost 1,000 linux local privilege escalations using our cyber vulnerability pipeline, at the cost about $1,000 per vulnerability. an emerging scalable approach to significantly hardening both open source and proprietary projects
33
29
8
707
84,560