Matt Turck

VC at @FirstMarkCap. Host: MAD Podcast; Organizer: Data Driven NYC, Author: MAD Landscape.

[Matt Turck @mattturck]

Inside Nemotron and NVIDIA's AI lab: my conversation with Bryan Catanzaro (@ctnzr).

@nvidia is a chip company. So why does it put hundreds of researchers on building AI models - and then give them away for free? We go deep into the Nemotron models, what it takes to build a top AI lab, and the future of frontier AI.

01:33 - Is open source AI catching the frontier?

05:29 - Do closed labs blocking distillation slow open source down?

07:42 - Is the US falling behind China?

10:30 - Why companies actually choose open models

12:39 - A "crazy" 2008 bet: machine learning on GPUs

15:33 - Working with Andrew Ng and Dario Amodei at Baidu

17:41 - Coming back to NVIDIA: DLSS and the birth of Megatron

21:55 - The real reason NVIDIA builds its own models

24:28 - Is Moore's Law really dead?

33:37 - The Nemotron family: Nano, Super, Ultra

35:09 - Built for agents: why NVIDIA bets on speed

36:02 - How you train a 550B model in 4 bits

39:25 - Hybrid Mamba-Transformer, explained simply

42:31 - Mixture of experts, and why NVIDIA built NVL72 around it

47:26 - Why a 1-million-token context window matters

49:26 - Multi-token prediction: how the model predicts 5 tokens at once

52:47 - Multi-teacher distillation: teaching one model from many

58:01 - Where reinforcement learning goes next

01:00:16 - Inside NVIDIA's research org: "the mission is the boss"

01:04:03 - How NVIDIA decides who gets the GPUs

01:10:53 - Why NVIDIA still feels entrepreneurial after 33 years

01:12:58 - Why Bryan doesn't believe in the singularity

01:17:50 - The AI backlash

01:19:18 - The controversial case: open AI is safer than closed

[Matt Turck @mattturck]

@tanayj

[Matt Turck @mattturck]

Tough game but France got it done. Between Argentina and this game today, refereeing is becoming a major issue in this tournament. 3 yellows for France and zero for Paraguay who played like thugs?

Also, France, stop passing the ball to Rabiot for long-range shots lol

[Matt Turck @mattturck]

@Patzerat9 @BorisBastide yeah that's fair

[Matt Turck @mattturck]

@jasonoliver yeah shocking honestly

[Matt Turck @mattturck]

@joshwaw97 last night was just ridiculous

[Matt Turck @mattturck]

@StuartBlitz that would tear my personal identity apart lol

[Matt Turck @mattturck]

@mattschaar 💯

[Matt Turck @mattturck]

Chills - Thierry Henry on growing up playing in the suburbs of Paris (and why 54 World Cup players are from there)

[Matt Turck @mattturck]

Little known fact, Morocco's goalkeeper Bono had a prior career in as lead singer of a rock band

[Matt Turck @mattturck]

@NickBForaTravel Awkwardly enough, the referees are sitting in the row behind me on the plane 😂

[Matt Turck @mattturck]

At Miami airport, seeing Cabo Verde fans who look like they cried all night. The reality of the World Cup is that 47 teams will go home heartbroken but you can tell that coming so agonizingly close, against the odds, crowd and referee hit hard. All a deeply human adventure.

[Matt Turck @mattturck]

and Vozinha, man, what a goalie. And that insane 2nd CBV goal? All of it in a crowd that was like 99% Argentinian. Just epic.

[Matt Turck @mattturck]

Best match of the World Cup so far. Cabo Verde ultimate Cinderella story, almost got a 3rd. Argentina is going to have to do *a lot* better going forward. #ARGxCBV.

[Matt Turck @mattturck]

@StuartBlitz goalie is amazing

[Matt Turck @mattturck]

@ctnzr This was truly awesome, thanks again for doing it Bryan

[Bryan Catanzaro @ctnzr]

Why are open technologies for AI so important?
How and why is NVIDIA building Nemotron?
What can we learn from China’s AI efforts?

Great conversation with Matt Turck.

https://t.co/GwNTd6If3b

[Ivan Burazin @ivanburazin]

Nobody can define what a sandbox is because the goalpost keeps moving.

The evolution of what agents need from a sandbox:

Stage 1 (code execution)

Your agent needs to run Python, analyze a CSV, and solve a math equation that ChatGPT can't do natively. So you spin up a tiny isolate

Stage 2 (coding agents)

Now the agent needs to clone a GitHub repo, edit code, install packages, run it, and preview the output. You need a full Linux machine

Stage 3 (agent lives inside the sandbox)

Now security matters. Can the agent see your tokens and credentials? What can it access on the internet? Do you need a firewall?

Stage 4 (RL workloads)

Now you want speed, throughput, concurrency, and spin up anywhere between 50,000 and 500,000 sandboxes simultaneously in seconds

Stage 5 (general-purpose knowledge work)

Legacy apps, internal tooling, workflows - everything lives in Windows. Linux sandboxes won't cut it here

The tools that serve Stage 1 don't fit into Stage 3. The ones that fit into Stage 3 don't work well in Stage 4. And so on. What your agent needed a sandbox for 18 months ago is completely different from what it needs today. And what it'll need in 6 months doesn't exist yet.