Richard Li

Entrepreneur. Advisor. AI and inference. Prev: ceo/founder @ambassadorlabs; product @duosec; corp dev/strategy @rapid7.

[Richard Li @rdli]

I've been building an #ai application for a little while now, and wrote up my 7 macro takeaways about building an AI app that I didn't know when I started.

https://t.co/mhv2ubrxnB

[Richard Li @rdli]

🔥

[Arthur Zucker @art_zucker]

A quick update on the future of the `transformers` library!

In order to provide a source of truth for all models, we are working with the rest of the ecosystem to make the modeling code the standard.

A joint effort with vLLM, LlamaCPP, SGLang, Mlx, Qwen, Glm, Unsloth, Axoloth, Deepspeed, IBM, Gemma, Llama, Deepseek, microsoft, nvidia, internLM, Llava, AllenAI, Cohere, TogetherAI.....

[Richard Li @rdli]

This is the agentic runtime stack—what’s needed to run autonomous, goal-directed agents in production.
#AgenticAI #LLM #AIinfrastructure

[Richard Li @rdli]

#Agents aren’t the future—they’re already here.
But building them? That takes a whole new stack.

🧠 AI
⚙️ Durable execution
🧱 Frameworks
🗂 Context
🛠 Actuators

Check out the breakdown (with @jflomenb and @Wing_VC):
🔗 https://t.co/gnk2BzYZeH

[Richard Li @rdli]

Context Is King 👑. Smarter agents need richer context—not just prompts. They aggregate context from multiple sources:

🧠 Knowledge (vector databases)
💾 Memory (short/long-term)
🌐 Actuators (APIs, sensors)

Check out the full Agentic Runtime Stack:
https://t.co/gnk2BzYZeH

[Richard Li @rdli]

APIs ≠ skills.

Today’s agents do more: write code, generate 3D models, browse the web.

We call these actuators—not just tools, but new capabilities.

With @jflomenb @wing_vc:
🔗 https://t.co/gnk2BzYZeH
#LLM #AgenticAI #AItools

[Akka @akka_io_]

Final chance to register for tomorrow's webinar with @InfoQ

Learn how to design and implement the next generation of AI-powered services with Tyler and @rdli.

See you there!
https://t.co/zcSap3fDKf

#InfoQ #Java #AI

[Richard Li @rdli]

Most apps act fast. Agents don’t.

They pause, retry, wait hours or days. That’s why they need durable execution—resilient workflows that persist across failures.

Check out our post on the agentic runtime stack:

https://t.co/gnk2BzYZeH

[Richard Li @rdli]

I've noticed this too! It's a total sycophant, which is not helpful if you're looking for more critical thinking.

[Sam Altman @sama]

the last couple of GPT-4o updates have made the personality too sycophant-y and annoying (even though there are some very good parts of it), and we are working on fixes asap, some today and some this week.

at some point will share our learnings from this, it's been interesting.

[Richard Li @rdli]

Agentic frameworks are expanding fast—most started with some sort of syntactic sugar around prompts, but they've since expanded into durable execution & memory.

I teamed up with @jflomenb + @wing_vc to map the stack 👇
🔗 https://t.co/gnk2BzYZeH
#LLMops #AgenticAI

[Casper Hansen @casper_hansen_]

2.1k stars, 2+ million downloads, and 7000+ models on Huggingface later, and I am officially ready to retire my long-time project AutoAWQ ⚡️

Proud to say that AutoAWQ has been adopted by the @vllm_project and will now be maintained by 55+ contributors 🥳

[Richard Li @rdli]

4. There can (and are) viable software inference businesses, but they need to be decoupled from GPU supply.

[Richard Li @rdli]

Fascinating! I have not thought deeply about the GPU market but clearly @evanjconrad has. GPU market is radically different from CPU market in a bunch of ways.

[Latent.Space @latentspacepod]

🆕 SF Compute: Commoditizing Compute

https://t.co/Wvedfmr4Ka

We're excited for our latest deep dive into the compute market with @evanjconrad of @sfcompute!

It should not be normal for the prices of one of the world’s most important resources right now to swing from $8 to $1 per hour (as @picocreator observed) based on drastically inelastic demand AND supply curves - from 3 year lock-in contracts to stupendously competitive over-ordering dynamics for NVIDIA allocations — especially with increasing baseline compute needed for even the simplest academic ML research and for new AI startups getting off the ground.

The entire point of SFC is creating liquidity between GPU owners and consumers and making it broadly tradable, even programmable. As we explore, these are the primitives that you can then use to create your own, high quality, custom GPU availability for your time and money budget, similar to how Amazon Spot Instances automated the selective buying of unused compute.

The ultimate end state of where all this is going is GPU that trade like other perishable, staple commodities of the world - oil, soybeans, milk. Because the contracts and markets are so well established, the price swings also are not nearly as drastic, and people can also start hedging and managing the risk of one of the biggest costs of their business, just like we have risk-managed commodities risks of all other sorts for centuries. As a former derivatives trader, you can bet that swyx doubleclicked on that…

Also to end off, we of course had to ask about how on earth SFCompute manages to have such immaculate vibes....

Timestamps

[00:00:05] Introductions
[00:00:12] Introduction of guest Evan Conrad from SF Compute
[00:00:12] CoreWeave Business Model Discussion
[00:05:37] CoreWeave as a Real Estate Business
[00:08:59] Interest Rate Risk and GPU Market Strategy Framework
[00:16:33] Why Together and DigitalOcean will lose money on their clusters
[00:20:37] SF Compute's AI Lab Origins
[00:25:49] Utilization Rates and Benefits of SF Compute Market Model
[00:30:00] H100 GPU Glut, Supply Chain Issues, and Future Demand Forecast
[00:34:00] P2P GPU networks
[00:36:50] Customer stories
[00:38:23] VC-Provided GPU Clusters and Credit Risk Arbitrage
[00:41:58] Market Pricing Dynamics and Preemptible GPU Pricing Model
[00:48:00] Future Plans for Financialization?
[00:52:59] Cluster auditing and quality control
[00:58:00] Futures Contracts for GPUs
[01:01:20] Branding and Aesthetic Choices Behind SF Compute
[01:06:30] Lessons from Previous Startups
[01:09:07] Hiring at SF Compute

[Richard Li @rdli]

3. GPU businesses such as CoreWeave should be viewed through more of a finance lens than a tech lens.

[Akka @akka_io_]

The era of #agenticAI is here.

#AIagents are replacing manual workflows—making decisions, taking action, scaling fast.

Learn how to design and implement the next generation of AI-powered services.

🔗 https://t.co/8w6AqvXBso
#Java @InfoQ @rdli

[Akka @akka_io_]

🚨 We're 1 day away! 🚨

Agentic AI is reshaping software—but scaling it isn’t easy.
→ TPS skyrockets
→ LLMs struggle with latency
→ Costs add up fast

How do you build services that can handle agentic scale without breaking the bank?

Join Tyler Jewell & @rdli tomorrow at 10 AM ET as they break it all down. Live Q&A included!
📅 Register here→ https://t.co/Jzpf59tfb6

#Akka #AppDevelopment #CloudComputing #CloudNative #DevOps #DistributedSystems #SoftwareDevelopment #AgenticAI #LLMs #LLM #AIAgent #AI

[Richard Li @rdli]

Join me, @TylerJewell, and the @akka_io_ team next week as I share some key lessons learned in building agentic #AI systems over the past year.

https://t.co/VA9YTBQM99

[Richard Li @rdli]

My friend @bnfb introduced me to a critical component of product requirements: setting a price you're willing to pay. This seemingly simple change creates focus, reduces risk, and improves communication. More:

https://t.co/bRbdNRBspH

[Richard Li @rdli]

@rekurencja @adpirz @rakyll I also think that strategies like model distillation are already reducing the amount of data plumbing work. Depending on what you’re trying to build, data may very well be less than 80%.