←
Cody Blakeney22,736 posts
Cody Blakeney avatar

Cody Blakeney

@code_star

Leading research at @arcee_ai | Formerly Data Research Lead @DbrxMosaicAI | Visiting Researcher @meta | Ph.D | #TXSTFOOTBALL fan | linktr.ee/code_star

◷ Joined August 2011▧ 2,125 Photos and videos
Cody Blakeney avatar
Cody Blakeney@code_star·Sep 16

I’m so incredibly proud of what we are building at @arcee_ai . It’s a truly special place. We are only getting started though. We’ve been quiet for a while, but we haven’t been idle. I can’t wait to share more about what we have been grinding on for the past ~6 months. I won’t spoil anything, but you can get some idea from this announcement.

Today, we are announcing our Series B funding round, valuing the company at more than $1B. This round accelerates our next-gen Trinity models across diverse infrastructure, expands our work with the DOE and national labs on Genesis-Science-1, and enables us to build the platform teams need to build, evaluate, deploy, and operate open models in production. We are grateful to our team, partners, open-source community, and investors. Led by @Vista_Equity, Cambium Capital, and @emergencecap, with participation from AI10 Ventures, @Hitachi, IAG, @M12vc, @p7ventures, and @Wipro.

Post attachment
Cody Blakeney avatar
Cody Blakeney@code_star·8h

I’m fairly certain this x years left to contribute to the world intellectually has always been a psyop by big labs for their best and brightest agi pilled employees to keep them working longer hours and discourage them from going to competitors. Take just a moment to consider what it means seriously. Why 1-5 years? (I’ve been hearing 1-5 years for about 2 years now fwiw). Why does contributing to the world intellectually even really matter to the extent that one must recenter one’s life around doing it in this window? Admittedly I started gradschool knowing full well I was going into industry on day one, and I told my advisor that before I started. That’s a bit unusual I admit. I like to say I didn’t know what a PhD was and accidentally became a scientist. But I did fall in love with the process, unraveling mysteries, and experimental design. The process of walking through mentally who could you take something abstract and making it into a testable hypothesis. What I will say is the problems don’t become more or less interesting because a computer can do them for you. If you are working on something you dislike so much you wouldn’t do it if there was another person that was faster or better at it, it’s not the faster better persons fault … you just don’t actually like doing it. Maybe it’s just a prestige thing idk, but sounds like if it was the case then it was never the “intellectual contribution” or joy of problem solving you really cared about in the first place. Don’t over think this shit. Just do what makes you happy. If AI solves that problem. Great, I promise there will be more out there waiting. Hell even if there is not this existential threat to your own lives meaning mindset isn’t helpful or healthy.

You might have 1–5 years to contribute intellectually to the world before AGI. What should you do with them? Excellent advice from @jaschasd (Anthropic) at Stanford. Skip Netflix tonight and watch this instead.

6:42
Arcee.ai avatar
⟳ Cody Blakeney retweeted
Arcee.ai@arcee_ai·20h

Huge thanks to @sgl_project for hosting and having us as a sponsor at the first SGLang Summit. Awesome group of people, and great conversations with companies across the AI ecosystem, with a free matcha bar to boot. See you next year. ✌️

2,000+ people. Two days. The first SGLang Summit. Together, we announced SGLang v1.0, released Miles v0.2, and launched the SGLang Consortium with nine founding members: @nvidia, @AMD, @intel, @Google, @Meta, @thinkymachines, @Cloudflare, @modal, and @radixark. SGLang is now used by 332 organizations. At Meta, it powers production inference for products serving billions of users. Teams from Modal, Cloudflare, LinkedIn, and others shared their production lessons. Across 90+ technical talks and six hands-on workshops, the community explored what comes next for inference, reinforcement learning, and AI agents. To everyone building, deploying, and contributing: thank you for making this possible. Explore the announcements, engineering lessons, and moments from SGLang Summit 2026 in the link below.

Post attachment
Cody Blakeney avatar
Cody Blakeney@code_star·17h

“computa bring up the traces. Show me the bubbles. Make no mistakes”

Cody Blakeney avatar
Cody Blakeney@code_star·18h

Have to read the tech report to find out if they used Farming sim for this

Researchers have used reinforcement learning to teach a 5-tonne John Deere skid-steer loader to perform complex acrobatic manoeuvres autonomously. Trained in simulation and transferred to a real machine, the system can balance on two wheels, navigate rough terrain, spin, and control its bucket before landing safely. The research demonstrates the potential of AI-driven automation for heavy construction and industrial machinery. Video Credit: @leggedrobotics / @ETH Paper: wevlv.co/4kltYys #engineering #technology

1:10
Cody Blakeney avatar
⟳ Cody Blakeney retweeted
Cody Blakeney@code_star·18h

One labs mind virus is another labs shareholder value > requires a ton of work & political capital to make & advocate for something that feels invisible when good, and push back against efforts that ruin it -- and eventually people burn out

Cody Blakeney avatar
Cody Blakeney@code_star·18h

Meta, google, and Microsoft never had shortages of talented smart people. That is necessary but not sufficient.

Cody Blakeney avatar
Cody Blakeney@code_star·18h

This. Model training failures (llama 4) have always been org failures. You have to get everyone aligned on making “good model” This has lots of obvious and less obvious consequences that run into challenges for how traditional tech engineering teams tend to organize themselves. It’s a constant struggle to avoid falling into the lower entropy modes. It starts with caring and having taste though. Both require constant effort and passion. Both must be fed.

good model behavior requires care for craft from the top craft is hard at bigger companies with so many cooks in the kitchen, where product execs obsess over a/b test metrics and research execs over eval numbers. it requires a ton of work & political capital to make & advocate for something that feels invisible when good, and push back against efforts that ruin it -- and eventually people burn out so then you end up with bad model behavior, and maybe a few people complain internally but then you launch anyway and then twitter blows up and then oh no we need a code red on model behavior but none of the structural problems go away

Cody Blakeney avatar
Cody Blakeney@code_star·20h

Am I the only one whose timeline is 20% people quote tweeting obvious bait misunderstanding of the plot of Dune with obvious straight forward explanations of the key themes of Dune? I’m not complaining, it just seems easier for everyone to read the books.

Matthew Leavitt avatar
⟳ Cody Blakeney retweeted
Matthew Leavitt@leavittron·20h

Modaic: Because Mosaic was already taken. Twice.

introducing modaic: tinker for decision models train decision models that hit the frontier for your task, then serve them fast. > prefill-optimized inference: 2.69× vLLM’s throughput and 1/3 the latency > training from $0.75/1M tokens > serverless inference from $0.04/1M tokens > API, MCP, and CLI for your coding agents gpus ate our marketing budget so rick and morty will explain try it here: modaic.dev

1:34
Cody Blakeney avatar
Cody Blakeney@code_star·Nov 17

Isn’t that exactly consistent with their results? > We change one thing: for the first part of training, both models read a perturbed version of the text (for example, with noise added to the token inputs), and the student matches the teacher’s predictions on that. Then we switch back to clean text. I don’t think it’s fair at all to say most people distill with pretrained students, unless you think meta, google, and others are not representative of real world distillation.

So... I reproduced this paper, and I have good news and bad news. Good news: the method actually does what it says. Training from random weights, perturbed-input distillation beats standard distillation at equal steps, by about the margin the paper reports. Bad news: the gain is specific to training the student from random weights. Once the student already has pretrained weights and you distill a specialized teacher into it, which is how most of us use distillation, the method is consistently worse than standard distillation. Caveat: Small scale. Pythia-14M student, 70M teacher, ~5M tokens

Post attachment
Cody Blakeney avatar
Cody Blakeney@code_star·Nov 16

Finally checking my DMs for the first time in like 2 months. My god I’ve missed some things lol. Sorry if anyone was trying to reach me.

Cody Blakeney avatar
Cody Blakeney@code_star·Nov 16

It’s great seeing a new lab trying to tackle real problems where AI meets the world. Unbelievable team. This is not a vague promises smoke and mirrors group. Really excited to see what comes out of here. Also Raj is a beast. In addition to being an amazing professor, he co-mentored many of my interns while we were at mosaic / databricks. They are all superstars now. If you want to learn from one of the goats and do cool work you should apply. (You know, after you apply for Arcee 😉)

Today we're introducing @noeri_ai, a new AI company whose name literally means "the plural of intelligence" with a $110m seed! Modern problems require modern solutions. In particular we care about (scientific) discovery - you'll hear from others what that else it means below! I def took quite a leap of faith when I left Nvidia a bit ago to help lead the research team here. It's been an honor to be on this journey with @yejinc @guestrin and the entire founding team! Thanks to our investors @MenloVentures @MadronaVentures @Wing_VC and many others. At UCSD over the last few years, I started becoming rather obsessed with the idea of discovery shaped problems: where you want to optimize for solving for max reward instead of average, where the horizons measure in months, where you're rate limited by physics, and where there's immediate real world impact if you get it right. I've spent a lot of time with scientists and in wet labs trying to figure out how to ~accelerate~ (SWE & white collar are imp ofc but if that's all AI manages to automate, that would be very sad) We've already been seeing massive improvements with our customers (and hope to share lots more in the coming weeks)! If this sounds like something you want to work with us on, join our wait-list by reaching out create@noeri.ai I've always said half of my motivation to go work is to hang out with fun smart people, the other half is to do cool stuff (and a hidden third half is to have nice sweets everyday). We've got all three and we're hiring all across the board - research, eng, product, ops, sales, etc! Join us noeri.ai/#careers

2:17