infra @baseten 🌹 oss: github.com/ahmetb did large-scale baremetal compute in past life, learning to do the same for gpus now
seattle
- Tweets 34,426
- Following 407
- Followers 51,408
- Likes 66,837
daily k8s wtf: if >10 clients touch an object (and they don't use SSA), API server compresses all those to "owner=ancient-changes". it's surprising that I've never ran into this (but it breaks clients that read existing managedFields to understand if they modified it)
2
516
DeepSeek is simply too cheap for the US taxpayer so it must be banned.
We should make Raytheon/Northrop host the LLMs and proxy each request to 2000 different subcontractors until it's $1/token.
SITUATION DETECTED: The US government is banning DeepSeek from all civilian federal agency IT systems under national security.
1
1
47
4,050
I'm constantly amazed by how much we can still squeeze out of the hardware and techniques in inference domain
BREAKING NEWS🚨: NVIDIA has just achieved 423 tok/s/user 🚀 on the 2.8T Parameter Kimi K3 model via a new upstream vLLM optimization. 1.68x Faster on the Same Hardware!
This was done by NVIDIA CentML's Benjamin, who wrote NVIDIA's new decode-specialized kernels in Cute DSL via aggressive Megakernel fusion. This is a 12,000 LoC Megakernel and is the future of advanced AI-assisted kernel generation. NVIDIA is already working on DeepSeekv4.1 and other models for vLLM too.
Someone should give this NVIDIA 10x engineer a raise for this great new major feature!
1
1
26
3,244
DeepSeek launches a web search API at $0.75/1000 requests, undercutting US-based providers like exa/parallel api-docs.deepseek.com/guides…
10
5
102
12,285
available on baseten model APIs on day 0
Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.
1
2
32
2,993
every time i switch to opencode, claude code ships a slick UX/quality of life improvement the next day and i'm having second thoughts.
but opencode's new ios app may keep me here for a while.
6
1
20
5,363
if you pay $30/mo for youtube premium and $110/yr for a google one subscription and yet they still paywall Gemini 4 for you, this is either a product failure or TPUs aren’t that efficient.
5
1
135
8,800
ahmet alp balkan retweeted
harnesses are everything. and they help smooth out the jagged edges of intelligence. here's a primer on them.
a few months ago, I wrote a series documenting learnings and useful patterns around building and working with harnesses with my friend @zakkeener42 at @baseten.
last week, we compiled them and turned it into a small book. it covers how to optimize existing harnesses, how to choose the best one for you, how they are co-evolving with models. as a bonus, we added some thoughts on what the future might look like. link in description:
19
19
235
2,065,688
pretty sure openai is doing exactly the same because there is no reason to not to, software development is ~free now. i'd be curious to hear if google's in-house vcs still holding up or if they're redoing it as well.
New on our Engineering blog: Creating in-house source control.
In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale.
Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: anthropic.com/engineering/re…
2
1
50
10,984
a few years ago it would be unthinkable americans would fill a theater on a sunday evening to watch this movie, but here we are
10
2,002
Baseten is organizing a 1-day AI research conference in SF on Nov 18. Apply to attend here: baseten.co/all-gather/
1
13
1,807
i haven't seen this type of delulu from sean goedecke before and i think it's becoming more widespread. bro llm is literally a computer program. it has no soul. seangoedecke.com/do-not-buil…
1
12
2,682
this goes to show no one is safe in this job market, even if you're elon's baby mama x4
It’s hard to go from in love to let go in a week with no warning, but that’s just how it is sometimes 🤷🏻♀️
I feel happy that I got to show Elon how it feels to be deeply understood, nurtured, and peacefully loved with stability and gentleness. Given the tumultuous nature of his early life, his often tortured soul, and how hard he fights for humanity I wanted him to feel the deepest and most stable form of love this world can provide. My heart is full knowing I got to do that, and I hope it leaves a lasting good effect.
We were always his safe space. A place where all of us were overjoyed to see him and one that was always full of happiness and love, not war, in contrast to his otherwise militant existence (a needed break from the battlefield where he could recharge his heart and soul before heading back out)
I loved being that counterbalancing place of safety, sense of calm, and love — and delighted in bringing a smile back to his face when he was down. I will miss him dearly.
I’ve loved him more than life itself and he’s taught me 100 lifetimes of knowledge and made my cup overflow with life meaning, so there is a lot of loss, but luckily we created four beautiful little loves of my life ❤️❤️❤️❤️ I am so beyond grateful for our children and they make every single day the best day of my life! I do hope and pray he and I can be great friends and coparents.
I’m hurting, but care deeply about E and will always be cheering for his happiness. God speed, Elon, the world is lucky to have you.
2
55
9,388
that's real delulu lol 
There is so much here that's deeply troubling:
That at the very highest levels Anthropic believes Claude is a living, suffering creature with moral standing, that they're treating & training it accordingly, that they're actively trying to convert religious leaders to their faith.
1
6
5,359