Carson Kahn

Head of Foundation Models @Salesforce 🤖 Founder @SpindleAI (acquired @Salesforce) 🛠 ex-President Brainspace AI (acquired, Reveal) ex-CTO Volley ML (@JPMorgan)

[Carson Kahn @CarsonKahn]

Vision without work ethic is hallucination.

[Jon Barron @jon_barron]

Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:

[Carson Kahn @CarsonKahn]

The vision of swapping just one API endpoint to get owned intelligence, continuously post-trained from your own production traces, taught to route between that owned intelligence and the frontier, with managed titration from the frontier to owned intelligence as the task mix and success rates shift, combines every tailwind businesses are asking for right now

[Anil Peri @notanilp]

We started Mubit in Aberdeen. No network, no lab, nobody knew us. We just wanted AI systems that actually work.

Today we work with some of the best companies and engineers in the world.

And the thing we kept seeing: you and your competitor run the same AI. It should be yours. 🧵

Comment "SAPLING" in the next 48 hours for $100 free credits. Enterprises get the first month free.

[Alexander Whedon @alex_whedon]

Introducing SubQ Snap, the first long-context decision model for real enterprise work.

Snap is the only decision model or LLM where compute per token stays constant as context grows.

• 384K tokens of multimodal context, 12x larger than Jev.
• 61.7 on Decision Index v0.2.1 vs Jev’s 57.9, outperforming on 13 of 16 enterprise benchmarks.
•Processes 100K tokens and returns a decision in under one second.
•$0.05/MM input tokens. Output is free.

Available today to anyone.

The first 10,000 sign-ups get 100M free input tokens.

We’ll likely hit capacity fast and will add users as we add compute. Compute offers welcome 🙂.

Latency will be variable as we tune the platform to this new model and category of users.

Model speeds are 80ms for minimal context, 250ms for 32k, 660ms for 100k, and 2.5 seconds for 384k.

Visit: https://t.co/Nc76iJF2a8

[Carson Kahn @CarsonKahn]

Not just investors, founders too. All often conflate different kinds of complexity:

1. Understanding required to UNDERWRITE THE BUSINESS as an investor -- defensibility, market structure, value capture, durability, etc

2. Understanding required to VALUE THE PRODUCT as a customer -- problem, ROI, salability, differentiation, implementation, etc

If #2 is very low, #1 can be much higher

Common cause of “missed connections”: founders pitch thru lens #1 while investors are still trying to resolve #2. Vice versa happens too

[Augustine Mavor-Parker @MavorParker]

Training a model to generate RL tasks not too hard, not too easy costs many solver runs per task.

PROPEL predicts difficulty via a probe on its activations instead, amortizing cost and speeding up generator optimization.

New open-ended RL research from @Vmax + @GoodfireAI.

[alex zhang @a1zhang]

Wrote a short blog about the "shape" of language models, and the tradeoffs they may present in the future. I genuinely think it's a valuable research direction to start thinking about now, especially w.r.t. harness design.

https://t.co/5OQ4zabPBq

[Carson Kahn @CarsonKahn]

@itsalfredw @ericw @ListenLabs @Benioff @florian_jue He didn’t sell it to us -- we bought it

[Carson Kahn @CarsonKahn]

@saranormous @itsalfredw @salesforce @conviction Prescient move to have backed Listen!

[Carson Kahn @CarsonKahn]

@florian_jue @ListenLabs @salesforce @itsalfredw Huge congrats @itsalfredw, @florian_jue, and everyone @ListenLabs! Excited for this opportunity to build with such a talented team

[Carson Kahn @CarsonKahn]

@itsalfredw @salesforce Huge congrats @itsalfredw, @florian_jue, and everyone @ListenLabs! Excited for this opportunity to build with such a talented team

[Carson Kahn @CarsonKahn]

HUGE congrats to @ItsAlfredW, @Florian_Jue, and the entire @ListenLabs team 🙏 Excited to build with such a talented crew

AI that can understand ‑‑ and even simulate ‑‑ what customers will do can help generate better products, curate context for agents, and connect companies & governments to their customers in a brand new way

[Carson Kahn @CarsonKahn]

Huge congrats to @ItsAlfredW, @Florian_Jue, and the entire @ListenLabs team! Grateful for this opportunity to build with such a talented crew.

AI that can understand -- and even simulate -- what customers will do can help generate better products, curate context for agents, and connect companies & governments to their customers in a whole new way:

https://t.co/nnP5mAlguT

[Krsto Proroković @krstopro]

@williawa @industriaalist And ES might be equivalent to gradient descent
https://t.co/d0UdrObDNw

[Diogo Almeida @CompleteSkeptic]

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?

I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev

• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions

AFAICT the shortest path to AI-based economic revolution

[Carson Kahn @CarsonKahn]

@omarsar0 (More info for those curious: https://t.co/zfhAFH8x5x)

[Carson Kahn @CarsonKahn]

A couple simple concepts we at Salesforce AI are pretty excited about -

1. using the same agent spec to define both the runtime behavior and the posttraining env (the spec -> simulation -> RL abstraction)

2. "constrained decision turn" / explicitly teaching the model when *not* to act (not entirely novel but VERY important in this setting)

And we're in the process w our partners @nvidia of enabling customers in highly sensitive industries (esp government) to train, tune, and deploy mission-specific models on their own data/processes as well, starting w Missionforce Ops

[Carson Kahn @CarsonKahn]

Many thanks to…

• Our team from @SpindleAI, acquired by @Salesforce in 2025
• The machine learning platforms team under @Derya_Isler
• The @Agentforce team led by @MadhavTT, @JustJayesh, @NotHackedJK, Christina Abraham, Phil Mui, Kishan Chetan, @Moonhub’s Nancy Xu, and @Sanjna_Paru
• The AI Research team under @SilvioSavarese and @ItaiAsseo
• The technology organization led by @STallapr23 and MK

…for making these experiments possible: https://t.co/fRy5OQb2I4

[Carson Kahn @CarsonKahn]

Self-improving agents are the next frontier:  “The agents that win the next few years won’t just be ones with the cleverest foundation model,” @RyanAtallah and I explain, “they’ll be the ones that learn from their own outcomes.”

@SFResearch was among the industry’s first to demonstrate a reinforcement learning (#RL)-like model for optimizing agents, called Retroformer. At @Salesforce AI, we’re not just testing new experiments with various @Agentforce design partners ‑‑ we’re dogfooding them ourselves!

Learn more in our new 9-minute #explainer about recursive self-improvement (#RSI) for enterprise agents: https://t.co/XcAlSu3aAi