For decades, 2 things kept the internet safe: writing code was hard, and finding bugs was hard.
AI just ended both of those.
@Mozilla CTO @raffi laid it out in his guest essay in The New York Times called "The End of the Internet as We Know It," and this morning he takes the @AICouncilConf stage w/ me to talk about what comes next.
See you at 9AM!
Over the weekend, I reviewed slide decks from all the speakers in the Databases and Data Engineering track at the @AICouncilConf . The talks are incredibly strong across a wide range of topics — transactional, analytical, search, vector databases; ETL/CDC; real-world case studies at massive scale, all with a forward-looking perspective on the critical role of Data Infra in supporting AI.
I’m so excited about this track and expecting it to be absolutely epic!!
Huge thanks to all the speakers for being an integral part of it and helping make it happen Hannes Mühleisen from @duckdblabs , @nikhilbenesch from @turbopuffer , @J_ and Pierre Lacave from @datadoghq, @iskakaushik from @ClickHouseDB , @kelvich from @databricks , Bhargavi Reddy Dokuru from @netflix , Robin Tang from @artie_labs 🙏
The track is on on May 12 (Day 1), starting at 10 AM, don’t miss it if you’re attending AI Council.
@petesoder@ZeroPrimeVC
Most teams building AI agents are reaching for bigger context windows. @thesephist thinks they're solving the wrong problem.
Hear more from Linus, Head of AI at @thrivecapital, at @AICouncilConf next week — and get a preview in our Q&A here:
https://t.co/sDEaft0ygY
"Most importantly, how are we going to [build AI] in a responsible way?" — @EnoReyes, CEO of @FactoryAI
That's the importance of getting world-class builders in the same room. See you all next week.
https://t.co/EPmcuuOFr1
A lot of AI observability tools are fantastic as long as the security team never logs in. With @honeyhiveai v2 (announced today!) you can keep full raw traces inside your own environment and still look your CISO in the eye.
@mohak__sharma, @ds3638 and team have rebuilt HoneyHive so raw agent traces stay in a customer‑controlled data plane and evaluators execute on that data, w/ the control plane pared back to metadata & RBAC/rollout aligned to how big orgs divide teams and workloads.
It’s an important building block for running AI agents as first‑class auditable production systems in large, regulated enterprises. Congrats to the HoneyHive team on v2 - and see you next week at @AICouncilConf!
More on HoneyHive v2 from CEO Mohak, rolling out to users the next few weeks: https://t.co/PxmFsZeVbz
1 billion tokens in, 1 billion tokens out. Opus 4.6 runs you about $30,000, real-time. DeepSeek‑V4‑Pro async on @Doubleword_ lands closer to $4,100.
Roughly the same intelligence, ~86% cheaper. That delta is what @MeryemArik9 has been building around.
Most inference stacks were designed for humans sitting and waiting on a response - ChatGPT, Claude, Perplexity, Cursor, Codex. Everything optimized for that near real-time loop, including the spinner verbs you read while you wait.
An async agent is a different pattern entirely. It chugs along for hours and nobody's watching. What matters is the total cost when the job finishes.
The teams not lighting cash on fire are getting deliberate about which tokens need a frontier model and an immediate response. Sometimes you pay for the realtime closed frontier reasoner. Sometimes the async open model gets you there just fine.
This is the territory Meryem is covering at AI Council - long-running async agents that don't torch your token budget.
Her talk will cover strategies builders can use to maximize async agent performance while keeping inference costs under control, covering context engineering, compaction, cache maintenance, model routing and batch inference.
Highly relevant for builders.
@AICouncilConf 2026. May 12–14. See you in SF!
Earlier this month, a 30-person US open-source startup shipped a 400B-parameter Mixture-of-Experts reasoning model for long-horizon agents.
The model is Trinity-Large-Thinking from Arcee AI, built on Trinity Large, one of the most ambitious open foundation models ever trained from scratch by a US team. Its predecessor, Trinity-Large-Preview, is already one of the most-used open-weight models on OpenRouter.
OpenRouter, where Trinity has been racking up that traffic, is the unified API for hundreds of open and closed models. The inference runs on platforms like Fireworks AI, which serves open-weight and fine-tuned models to Cursor, Notion, DoorDash, and Uber. On top, you get applications like Kilo Code, the fastest-growing open-source coding agent. Four companies, four layers of the open stack — model, routing, inference, agent.
We've invited all four founders on stage at @AICouncilConf this year to talk about this "open layer."
Their upcoming talk, "The Open Layer: How Open Models, Routing, and Inference Are Reshaping Agentic Engineering," gets into what "open" actually means in 2026 (and where it falls short), when open-weight models win (and when they don't), and what it really takes to keep always-on agents reliable on this stack.
On stage:
@MarkMcQuade, Founder & CEO — @arcee_ai@cclark, Co-Founder & COO — @OpenRouter@dzhulgakov, Co-Founder & CTO — @FireworksAI_HQ@s_breitenother, Co-Founder & CEO — @kilocode
Four exceptional founders on one stage. Looking forward to this discussion!
May 12–14, SF.
https://t.co/JgiORdxjXN
Meet @lloydtabb. Bike mechanic, co-creator of Malloy, founder and former CTO of Looker.
At last year’s AI Council (fka Data Council), you could listen to Lloyd demo Malloy while making the case that semantic modeling is what makes LLMs actually useful on top of your data. Then you’d stick around for Office Hours to ask him a question.
Every year, people tell me the speaker Office Hours is their favorite part of the conference. This year will be no different. It’s not often you can be in the same room with your AI & data heroes and the builders of your favorite tools for intimate, small-group chats.
The @AICouncilConf 2026 agenda is live. Pre-plan your must-see talks and Office Hours to get the most out of the conference: https://t.co/46RjNwmc2I
Talked to @changhiskhan of @lancedb and it's got me thinking:
Most of the current data stack was built for a human hitting "search" a few times a minute. Not for an agent firing a hundred queries in parallel and chaining them across a long reasoning path.
Curious what others are seeing. If you're running AI in production right now — what's breaking first? Throughput? Latency? The coordination tax between systems?
Full conversation from our sit-down ahead of @AICouncilConf:
https://t.co/TM6XCbJkrU