Pete Soderling @petesoder

Open models have been gaining on leaderboards for a few months now but an analytics platform putting an open model in their product for the first time carries a bit more weight. Kimi K2.7 is now available in @_hex_tech, hosted on US-based infrastructure, at an intelligence level comparable to Opus and GPT-5.5: "An open model can be as cheap and fast as it wants, but speed means nothing if it confidently hands you the wrong number [...] We feel that with Kimi K2.7, we finally have an open model that’s on par with Opus and GPT-5.5 at data tasks." As they continue to improve, open models will leech usage off the proprietary players. Not overnight, because the market's growing too fast for anyone to notice at first. But it's started, and this is what it looks like. hex.tech/blog/kimi-in-h…

.@PrismML took Qwen3.6 27B (~54GB at FP16 ) and crushed it to 1.71 bits (5.9GB) or 1.125 bits (3.9GB) while keeping 95%/90% of benchmark quality. Math and coding barely move (95.3→93.4, 88.7→86.0) and agentic/tool-calling drops only 80→74 and vision drops only 72.6→65.2 in the 5.9GB 1.71-bit version. Why this is important: AI workloads are increasingly agent workloads, and right now that looks like 100s of context-accumulating calls sent to the cloud for any given agentic request. Prism is asking, "What if some of that work could be done quickly (and for free) on device?" From their blog post: "When a model capable of sustained agentic work fits on the device, the agent can live inside the product: the marginal cost of a hundred-step loop is zero, and the user's data never leaves the machine. Entire categories open up — persistent on-device agents, assistants that work offline, assistants that reason over private local data by construction. What has been missing is a model small enough to deploy this way and capable enough to trust with the work." Exciting release from PrismML! If you're curious to understand some of the techniques behind the accomplishment, @HessianFree (co-founder and co-head of research at PrismML) gave a fantastic technical talk about extreme quantization at this year's @AICouncilConf: 📺 Making Neural Networks Smaller: Quantization and Pruning youtube.com/watch?v=2qeEal…

Congratulations @cusp_ai! Chad Edwards and @wellingmax announced a $450M Series B today at a $2.6B valuation. They also launched the AI Materials Foundry, a global network of 45+ founding members with NVIDIA providing the compute and Meta's FAIR team contributing frontier atomistic chemistry models. Incredible progress in a short period of time towards realizing the “search engine for new materials.” I'm excited but not surprised by the quick pace. Speed is a big part of their product. For customer Kemira, Cusp screened a search space of 300 trillion molecular structures and delivered 20 validated candidates in 6 months. This kind of screening had previously taken Kemira years. The AI Materials Foundry now points that speed at 45+ partners working across semiconductors, clean energy and advanced manufacturing. Proud to be investors. Congrats again to the whole CuspAI team.

Here is @schrockn, founder of @dagster, on stage at @AICouncilConf in 2019. Cross out "big data," write "AI," and it's a headline you've seen a dozen times this year. The reason hasn't changed much: models get the attention, but the data decides whether the models actually do something useful. We just call it "context" now. What's more, the declarative pattern Dagster was built around is now the one winning in agent work: define the outcome up front, let the model verify against it, and loop until it passes. Declare what should exist, then prove it got produced. This is what makes @PrefectIO acquisition of Dagster click: Dagster defines the outcome, Prefect runs the work, FastMCP handles access. One stack for making automated work trustworthy. Huge congrats to Nick, @floydophone, and the Dagster team!

You don't just turn on a vacuum and hoover up all your company data and call it a context layer. People get things wrong in slack. People get things wrong in email and zoom. And when presented with an answer based on a stale or incorrect context, users often don’t know the answer is wrong. An effective context layer is closely managed with all the maintenance and care of a garden. Artifacts are watered, weeded, pruned, fertilized. The knowledge base is carefully curated and managed. "The model decides how smart your agents are. The context layer decides whether they're right." Well said, @arudis linkedin.com/posts/swaroopj…

.@ashpreetbedi is living in the future. linkedin.com/posts/ashpreet…

"Waiting for perfect" is the worst mistake in AI and data right now. It's the kind of mistake that aligns you with the likes of Dilbert's boss, Lumberg, and other hilariously terrible caricatures of corporate bureaucracy. I asked @Snowflake's Bala Kasiviswanathan what leaders are getting wrong around AI and data right now and this was his answer. He said leaders are getting stuck waiting for perfect. And while they wait, they do the one thing that guarantees they lose: they don't let their teams try anything. The great ones are doing the opposite right now: building platforms safe enough where nobody can hurt themselves or the company, and then letting everyone loose to learn. Here's what "safe" means, and why it works: → The agent runs as you. Snowflake's Coco - as an example - acts on your behalf, so it inherits your roles and permissions. It can't run a destructive operation on a database that your own access wouldn't allow. Coco or not, the same rules hold. → A human stays in the loop. The agent can't touch the risky stuff on its own; someone has to say yes first. → Every move is logged. Audit trails tell you who did what so you can learn from it, and catch anyone acting in bad faith. Get those 3 right and the risk of letting people experiment drops to almost nothing. Stop hunting for perfect, build the safe playground and let your people play. 🎧 My full talk with Bala is on the @ZeroPrimeVC podcast: zeroprimepodcast.com/snowflakes-bal…

@TheEthanDing @TextQL @ZeroPrimeVC Full episode here: youtube.com/watch?v=PwhlxJ…

Own your storage. Own your business logic. Own your compute. If you're renting all three, bound to one platform, you won't scale in the agent era, says @TheEthanDing, Co-founder & CEO of @TextQL. We get into it in our latest @ZeroPrimeVC episode.

@BEBischof @Theoryvc @AICouncilConf @lancedb @motherduck Oh no. There will be more!

Enterprise data doesn't need to live in one place anymore with tools like @Snowflake's Coco, says Bala Kasiviswanathan, VP of Dev & AI Experiences at Snowflake. If data access isn't the constraint anymore, what should data leaders focus on instead? We get into it on @zeroprime.

finance teams watching devs use fable to center a div

@Theoryvc @motherduck @lancedb @AICouncilConf Amazing

@davemcclure @AICouncilConf Thank you Dave - have learned so much from you 👊

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!

We are gonna have some fun 😈

AI Council @AICouncilConf

2 months ago

Our Day 02 afternoon keynote "Beyond the API" is beyond stacked. New panelist just added... welcome back @charles_irl - Member of Technical Staff @modal, joining: @BEBischof, Head of AI, @Theoryvc Charles Zedlewski, CPO, @togethercompute @tuomars, Research Scientist, @nvidia

0 0 5 5K 0

It me

AI Council @AICouncilConf

2 months ago

You: at AI Council workshops all day Them, admiring your extraordinary diligence:

1 0 3 575 0

I'll need a green room and a bowl of M&Ms (no brown ones)

AI Council @AICouncilConf

2 months ago

Starting a petition for a @petesoder Day 02 encore.

0 0 1 218 0