kevin yang

rare token prospector. curr: in sf/nyc building. prev: cs/ee @stanford

[Ronak Malde @rronak_]

🏹 Day 3 of the 5 Days of Trajectory!

We are open sourcing a training stack for continual learning, in collaboration with SkyRL (@NovaSkyAI) and Anyscale (@anyscalecompute)

At Trajectory, our mission is to bring the capability of continual learning to every team and company.

Our contribution today is a multi-tenant, continual LoRA (C-LoRA) training stack that is built for workloads that are repeatedly spinning up and down.

Links to get started below!

[kevin yang @YangKevinG]

@emilyzsh @ryanjdaniels @jsarihan you're the goat emily! keep shining

[kevin yang @YangKevinG]

@philhchen @tkkong Congrats Phil!

[kevin yang @YangKevinG]

Congrats Sri + team! from startup garage together in fall quarter 2020 to now -- continuously inspired by your humility, tenacity, and insight. Exciting to see how far Latent has come, the future is bright for AI in healthcare

[Latent @latent_health]

Excited to share that Latent has raised $80M to build the clinical reasoning engine that closes the gap between diagnosis and treatment.

This round is co-led by @sparkcapital and @transformcptl, with participation from @Conviction, @MCK_Ventures, @generalcatalyst, and @ycombinator.

For the first time, AI makes it possible to reason through patient data, interpret drug criteria, extract key evidence, and orchestrate clinical workflows at scale. Latent is that reasoning layer.

Today, over 45 of the top U.S. health systems, including Yale New Haven Health, UCSF Health, UCLA Health, Mount Sinai Health System, and Vanderbilt University Medical System all use Latent to perform high-stakes clinical knowledge work.

We've helped over 2 million patients access life-saving medications faster and reduced denials by more than 30%.

We're expanding our clinical reasoning engine across every process where clinical knowledge must be translated into action, and building a team to match the scale of the problem.

[kevin yang @YangKevinG]

Congrats on the launch @EugenAlpeza!

[Eugen @EugenAlpeza]

We’re out of stealth.

Today, we’re also announcing our Series A led by @sequoia , @8vc , and @A_StarVC , bringing our total funding to $30M+.

Every enterprise needs to teach their AI how to do work. We build agents that reverse engineer enterprise processes, then run them.

Read about the future of learning in the enterprise: https://t.co/ONBjA7MEEJ

[Kimi.ai @Kimi_Moonshot]

Introducing 𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 𝑹𝒆𝒔𝒊𝒅𝒖𝒂𝒍𝒔: Rethinking depth-wise aggregation.

Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth, we introduce Attention Residuals, replacing standard depth-wise recurrence with learned, input-dependent attention over preceding layers.

🔹 Enables networks to selectively retrieve past representations, naturally mitigating dilution and hidden-state growth.
🔹 Introduces Block AttnRes, partitioning layers into compressed blocks to make cross-layer attention practical at scale.
🔹 Serves as an efficient drop-in replacement, demonstrating a 1.25x compute advantage with negligible (<2%) inference latency overhead.
🔹 Validated on the Kimi Linear architecture (48B total, 3B activated parameters), delivering consistent downstream performance gains.

🔗Full report:
https://t.co/u3EHICG05h

[kevin yang @YangKevinG]

that boy tuff

[Mason Wang @masonwang025]

lead author's 16 btw
(incredible work @nathancgy4 😁)

[kevin yang @YangKevinG]

@milichab @SpaceX @xai @JasonBud Lfg, congrats Andrew!

[kevin yang @YangKevinG]

@masonwang025 back in my day all we had was rubbery chicken, wtf man

[kevin yang @YangKevinG]

congrats on the launch @ryanbrewer - exciting to see you bring this to life!

[Ryan Brewer @ryanbrewer]

Excited to share that I’ve joined OpenAI to help build the future of financial intelligence.

My focus is connecting models directly to the data sources, tools, and workflows analysts use every day. My team’s first product is the Excel plugin.

After software engineering, finance will see the benefits of model improvements more acutely than almost any other field.

This is just the start for us! More to come soon!

[Nathan Hu @NathanHu12]

What does reasoning fine-tuning actually change inside a model?
In our new paper, we introduce transcoder adapters to learn sparse, interpretable approximations of how reasoning fine-tuning changes MLP computation. 🧵

[kevin yang @YangKevinG]

study model visual reasoning

[Nicolai Ouporov @nicoup]

Excited to share new benchmarking work from @fleet_ai &amp; friends.

We challenge frontier models to draw!

Surprising, across the entire frontier, models are really bad. The ways they fail can teach us about how AI perceives our world 🧵

[kevin yang @YangKevinG]

alas DSL was misguided here, turns out LLMs can be strong enough at any language. DSLs were not bitter-lesson-pilled enough in retrospect (many things of the 2023 era were not)

[kevin yang @YangKevinG]

@nicoup seems we have been on the same page about things for a while :)

[Nicolai Ouporov @nicoup]

You gotta respect venture capitalists. Just absolute love of the game

[Michelle Lim @michlimlim]

When the cost of code goes to zero, marketing is your only advantage.

Introducing Flint. It builds you a unique page for every ad, keyword, and customer.

We’re already doubling conversions for @Cognition and @Graphite. Sign up at @tryflint.

[Rhythm Garg @rhythmrg]

It will be increasingly important for enterprises to have crisp, automated definitions of what is "good" and "bad" when building agents for a task. These are what are ultimately used as the north star when improving agents (both via prompting and training) on the task.

DoorDash is leading the charge here; we worked with them and their internal human experts to build an automated grader for an important workflow in merchant onboarding, and then used that artifact to build a useful agent together.

[kevin yang @YangKevinG]

@joon_s_pk @karpathy @drfeifei @adamdangelo @rauchg @scottbelsky Congrats on the launch, @joon_s_pk! Long-horizon world simulations are the future.

[Nicolai Ouporov @nicoup]

Deep Think is an absolute feat.

AI models are starting to pour out of the digital world and confront reality. Just as AI scientists are now controlling labs, we will soon see AI manufacturers designing parts and running autonomous factories. Proud for @fleet_ai to play a role in this shift

[kevin yang @YangKevinG]

@flappyairplanes very exciting, congrats Ben + team!