Video · 2026-08-06 · 46m · 6 moments

How Open Source Became AI's Backbone | Inferact with a16z

✦ AI generated

timeline · colored by role

Claim◆

About a year ago, open source became genuinely central to AI because application companies realized they could not build on closed source alone — they needed their own mid-training, post-training, and inference control that only open source enables.

Matt argues that while closed-source models remain more widely used overall, the field passed a threshold roughly a year ago when leading application startups (Cursor, Decagon, Harvey) concluded they had to do their own training and deployment tricks — which requires building on open source rather than closed APIs.

transcript

Matt Bournestein: I do think we passed a threshold in like I want to say about a year ago where a bunch of smaller companies or like new application companies as they were trying to figure out how do I really build an AI without just being a wrapper on top of open AI. The answer to that question turned out to be open source. I mean this is what cursor did... and and a bunch of other like really really strong application level startups sort of made the determination we can't build just on closed source. We need to do our own mid-training our own post- training our own sort of inference and deployment tricks and all of that means it must be built on top of open source...

Definition

VLM is an inference engine whose job is to turn available GPUs into a running endpoint for intelligence, making it critical software — like databases and operating systems — that powers how people build and run AI models.

Simon introduces what VLM is and where it sits in the AI stack: it is an inference engine that turns any available GPU into a running endpoint for intelligence, analogous to databases and operating systems as critical software powering the AI economy.

transcript

Simon Mo: VLM is a inference engine. That means its job is to turn available GPUs into a running endpoint for intelligence. So that means it is kind of like databases and operating system other critical software to power uh this uh economy or power of the AGI that everybody really uses today to ensure they can have uh cost effectiveness, efficiency, reliability and also always staying on the frontier because for VRM we support more than a thousand model architecture up to today and a lot of those are proprietary But also a lot of those are openweight right...

Mechanism

Open-weight AI models need economic incentives and funding mechanisms to be sustainable, because unlike classic open-source software you cannot train a frontier model through donated volunteer time — it requires millions or billions of dollars of compute.

Matt argues the economics of open-weight models differ fundamentally from open-source software: you cannot volunteer to train a frontier model for fun since it needs billions in compute. He adds this is even more pressing for Chinese labs, where without commercial funding the alternative source is government money.

transcript

Matt Bournestein: I I I don't think it's greed at least what I've seen from open source model labs... Open source models really what we're talking about are open weights... And it's just not software, right? Like an AI model is not software at the end of the day. Um, and so open- source software used to be supported by people donating their time or big companies kind of authorizing their employees to donate their time... That really doesn't work in AI... Like I can't just like go home at night and like train a frontier opensource model with friends for fun. Like we need millions or billions of dollars of computing resources in order to do it...

gives example · 1

Claim◆

Using a closed proprietary model API means suffering the control and moderation tradeoffs, so for trusted use cases people will go with open-weight models by default because that is where they can control their own guardrails.

Simon argues that closed model APIs have arbitrary and difficult-to-enforce guardrails with frequent false positives, citing Hugging Face using a Chinese open-weight model to contain an attack and Infract developers switching from Claude to Kimi K3 after false-positive safety triggers. He predicts people will turn to open-weight models by default for trusted use cases.

transcript

Simon Mo: In the end is about all the closed proprietary model API. their guard rails are a little bit arbitrary but also very difficult to enforce... So if moderation is never solved which is going to be very very hard then there's always a place where you have a model where you know and trust that you are publishing to and be be able to use from... in the future also see for the trusted use case people will go to open way by default because that is where you know for sure that uh the guardrail is uh lessened or you can control your guardrail for trusted use cases...

rebuts · 1

Prediction

Frontier capability is no longer the differentiator between open-weight and closed-weight models — the real battleground is who builds the best environment for the model to improve itself, including data and learning loops.

Simon argues there is no big capability gap between open and closed frontier models today, because they come from the same ingredients — compute clusters, data, and brilliant researchers. The real differentiation over the next year will be who builds the best environments and algorithmic choices for models to improve, citing Kimi K3's front-end coding environment.

transcript

Simon Mo: in the end there's not much differentiation. is a more about the distribution strategy and go to market strategy and the capability wise I don't really see a big gap not even today because for how these model are coming to being they're really starting from the first principle... in the end there's not much... it's about who gets what data and then what are the environment you are building to let the model improve on itself... So the next year is all going to be about that is about how open way model labs are differentiating and really getting the model to meet the real world...

Claim

Open-weight research is both fundamentally different from and dependent on training-time environment and self-improvement processes, which means the 'distillation' narrative is overstated — progress comes from smart people with good algorithms and data, not from distilling another model.

Simon downplays the idea that distillation is central to Chinese open-weight labs' progress. He argues environment-heavy RL processes cannot be distilled, and what actually powers progress is smart people, interesting algorithms, and data — sentiments Matt extends into a policy point that turning off distillation would not solve anything since the labs are genuine innovators.

transcript

Simon Mo: I I I would lean to the latter part specifically as I mentioned previously environment matters so much today. So these are RO environments right these cannot be distilled like you don't have other people's environment to really distill a copy from... you cannot distill how the model learns within environment a lot of these are just not doable today... I really don't think from currently what we're seeing uh this is a big cornerstone of what's powering the progress today. In the end what's powering the progress is still just um really smart people with very interesting algorithms, data environment...

Highlight slides

Open Source Became Central to AI About a Year Ago✦ from: About a year ago, open source became genuinely central to AI because application companies realized they could not build on closed source alone — they needed their own mid-training, post-training, and inference control that only open source enables.Closed API Guardrails Are Arbitrary and Hard to Enforce✦ from: Using a closed proprietary model API means suffering the control and moderation tradeoffs, so for trusted use cases people will go with open-weight models by default because that is where they can control their own guardrails.Open-Weight Models Win Trusted Use Cases✦ from: Using a closed proprietary model API means suffering the control and moderation tradeoffs, so for trusted use cases people will go with open-weight models by default because that is where they can control their own guardrails.

Related episodes

Article · Interconnects

Kimi K3: The open-weights escalation

→

Video · Machine Learning Street Talk

Intelligence is collective, not artificial — Prof. Michael I. Jordan (UC Berkeley / Inria)

→

Video · The TWIML AI Podcast

AI Factories Explained: What’s Actually New?

→