VC at @FirstMarkCap. Host: MAD Podcast; Organizer: Data Driven NYC, Author: MAD Landscape.
New York
Joined February 2011
- Tweets20.9K
- Following2.5K
- Followers147K
- Likes67.8K
Pinned Tweet
Mistral is SO back: Inside Le Chonk, the West’s Biggest Open Model. My conversation with @PierreStock who runs the research organization at @MistralAI
0:00 Cold open
0:59 The Western open-weight comeback: Beam, Le Chonk, Muse, Nemotron
4:09 The US shutdown of Anthropic’s models - “AI is like electricity”
6:07 Why “Le Chonk”
7:56 “Mistral fell behind”: the comeback from the inside
10:29 Training on 3,800 GPUs in Mistral’s own datacenters
13:24 Meta keeps its best model closed, Mistral opens Chonk
15:40 Chonk vs Beam, Muse Spark Open and the Chinese models
18:44 Why Chinese labs are still at the frontier
21:08 Pierre’s story: FAIR, Yann LeCun, the “Mistral mafia”
25:14 From compressing networks to a trillion parameters
27:30 Mixture of experts explained: the hospital analogy
32:25 10-trillion-parameter models are coming
35:29 What Mistral borrowed from DeepSeek
36:46 One model, several reasoning modes
39:15 Native multimodality: satellites, manuals, computer use
41:56 Data, synthetic data and model collapse
46:32 Reinforcement learning as an exam
49:59 Where the efficiency comes from
52:18 Le Chonk is still training right now
53:39 What RL trains on: coding, cyber, finance, chips
58:13 Partial rewards and looping penalties
59:58 Did RL teach anything new?
1:03:16 The sandbox story: the model that pinged its trainer
1:07:44 Open weights vs cyber refusals
1:12:05 Timestral: time series on a CPU
1:14:18 AI doing AI research
1:16:22 Coming up: Mistral Large 5
Matt Turck retweeted
AI is making the late-stage IPO scramble a lot riskier.
Investors and auditors can dig into your numbers in minutes now, and if the answers sit in disconnected spreadsheets, it shows very quickly.
At @TechCrunch Disrupt, Sofia Fatakhova and I had a great roundtable discussion on how @Chime got its Finance team ready for life as a public company:
➡️ Her team made every number traceable to its source first, then moved seven separate models into one shared plan on Pigment, live months before the IPO.
➡️ Forecast accuracy got its own scorecard, with targets between 90 and 99%.
➡️ At least one week a month is protected for improving processes, with fewer meetings.
➡️ Every AI use case is tied to SOX controls and documented, so no one becomes a single point of failure.
Thank you to Sofia for sharing these learnings, to the TechCrunch team for having us, and to everyone who joined us!
Matt Turck retweeted
A small, on-device TTS models is competing with models up to 100× their size, and taking first place on specific benchmarks!
@GradiumAI's new Phonon model offers a concrete example.
On the Open TTS Leaderboard, Gradium reports:
- First place on intelligibility in German and Spanish.
- Top-three intelligibility in every language it evaluated.
- Leading CPU streaming results on time-to-first-audio.
These results also show why it’s useful to compare TTS models on more than a single overall ranking.
For the Open TTS Leaderboard, @EricBezzam separates several dimensions:
- Intelligibility, measured through word/character error rates.
- Generation throughput and time-to-first-audio.
- Predicted naturalness scores.
- Speaker similarity for voice cloning.
Leading on intelligibility in a particular language is a specific result, and so is leading on CPU response latency; but neither establishes that a model is best across every use case.
For on-device applications, Phonon’s combination of size, intelligibility, and CPU latency seems really relevant. For other (heavier?) workloads, language coverage, throughput, voice similarity, or delivery may be more important.
Leaderboard:
huggingface.co/spaces/hf-aud…
Gradium’s post: huggingface.co/blog/phonon-t…
Matt Turck retweeted
Our friends at @SurrealDB built a notes app on Phonon that analyzes your notes and reads the findings back to you, all on-device.
Phonon, our on-device TTS + SurrealDB Agent Memory.
Open source: github.com/surrealdb/surreal…
2028:
* Ossoff wins dem primary over Talarico (whose profile considerably increases throughout 2027) and AOC
* Rubio beats JD Vance (too tied to Trump) in the GOP primary
* Ossoff wins the presidency
(I'm a foreign-born VC, you can totally trust my U.S. political predictions)
Matt Turck retweeted
🤯
SurrealDB 4 is here. The database and memory for every agent. And with it, SurrealDB Earth. Serverless SurrealDB, built on @Cloudflare, that scales to as many agents and users as you need, wherever they are.
Start free on Earth, no card needed. 👉 sdb.li/4sHnxYH
Matt Turck retweeted
We've been working towards AI image and video detection for months, and now it's finally here.
There were a lot of challenges along the way. There are way more AI image/video models than frontier LLMs, and their outputs are more diverse. Most phone cameras, including iPhones, do some level of AI postprocessing to add detail to real photos.
We had to build a model that's robust to this postprocessing while understanding the difference between AI edits and fully AI-generated visuals.
I think the heatmap feature is a really important step forward and a huge differentiator over other models which typically report scores as a binary confidence.
Please try it out!!
As a customer, what I want: one central brain (AI) that can interact with me whether I text it or speak to it, whether I want to buy something or need support, and can actually do something helpful for me (agentic action) - 24/7, always in a good mood, instantly available.
The combination between @dialpad and @ada_cx creates a major new entrant pursuing this big vision.
As an investor since the Series A, excited for @mimurchison and team to go after it.
We are thrilled to share that @ada_cx and @dialpad are merging to create a new leader in agentic customer communications!
The companies of the future will be teams of people and AI agents, serving customers who increasingly have agents of their own.
Together, we’re going to power new experiences customers should expect, the collaboration businesses will need, and the communications that will make it possible.
Ada has served more than 1 billion people and deployed more than 500 customer-facing AI agents that handle complex interactions. Dialpad, founded by members of the Google Voice team, powers more than 50,000 businesses around the world.
As one company, we will make entirely new customer experiences possible: people and AI agents working as one team around every customer anticipating needs, taking on complex work and doing more than customers thought to ask. Together, we have the reach, resources and technology to bring that future to billions more people, faster.
Read Mike’s letter on what we’re building: ada.cx/full-stack-with-dialp…
Matt Turck retweeted
Serverless databases have a reputation for being the small, convenient option, not the fast one. I wanted to see how far I could push that.
This week we launched @SurrealDB Earth, serverless SurrealDB on @Cloudflare. It starts when a query arrives and scales to zero when nobody is using it. So I pointed it at 5 billion rows and asked it to count them.
In a single second, it read more than 4.5 billion of them. That is over 90% of the table, more than 350 GB of data, in one second. Every number here is from a real run, not a projection.
There was no cluster to provision and nothing to size. The parallelism appeared the moment the query needed it and was gone the moment it finished. And it is the full SurrealDB engine, with ACID transactions.
It is not perfect yet. 99.6% of the work finishes fast, but the slowest 0.4% still drags the full count out to several seconds. I know why, and my target is the entire query in under 2 seconds.
Honestly, I still find it hard to believe. I am convinced this design has orders of magnitude left in it, on reads, on writes, and in places I have not talked about yet. I will be recording a video soon with much bigger numbers.
The small, convenient option just got very fast.
I am really proud of what this team has built 🌍
The American democracy working as intended is a beautiful thing. It’s not that complicated: the majority of people don’t want either wokism or Trumpism. The system is keeping both in check, as it’s supposed to.
Matt Turck retweeted
Come take a look inside our Paris office.
@blwiertz from Tech: Europe chatted with our co-founders and the team about Gradium’s origin story, Kyutai, our friends next door, raising money as Europeans, and opening in San Francisco.
Stay for the voice design demo at the end.
youtu.be/-EDmwk2-S4c
Major moment for @SurrealDB
Building a database is an incredible grind - if you get that initial open-source pop, you then face years of unglamorous hard work to build a truly world class product
SurrealDB 4 is now that product. Multi-model (vector, graph, document), ACID compliant, w/ a rebuilt engine, native Postgres/MySQL support + massive scalability via @Cloudflare handling 10B+ peak rows/sec
The modern database for the agentic era.
SurrealDB 4 is here. The database and memory for every agent. And with it, SurrealDB Earth. Serverless SurrealDB, built on @Cloudflare, that scales to as many agents and users as you need, wherever they are.
Start free on Earth, no card needed. 👉 sdb.li/4sHnxYH
Matt Turck retweeted
Models will change. Enterprises need the freedom to adapt while maintaining control over their AI.
Proud to see Dataiku named a Leader in the IDC MarketScape: Worldwide Agentic AI Platforms 2026 Vendor Assessment.
dataiku.com/blog/idc-markets…
Hang it in the Louvre
We just crossed $100M in run rate.
Four quarters ago, we were at $1M. Then this year alone, we grew 37.5x.
Not long before that, we did a hard pivot- handed our old customers to competitors, rebuilt from scratch, and relaunched with a few months of runway left.
The bet was: models think on GPUs but work happens on CPUs.
So every agent that writes code, opens a browser, or runs an RL job needs a composable computer that's isolated, stateful, and ready in milliseconds.
$1M to $100M in about a year puts us in the same cohort as the likes of Anthropic and Cursor. It’s humbling to share that lane with companies we’ve learned so much from.