Mixpanel's founder, now running an open-model inference company, on when the frontier model is worth the bill and why every job is turning into code.
Posted
yesterday
Duration
Format
Interview
educational
Views
1.6K
43 likes
57 · 43
Big Idea
The argument in one line.
For most real work the frontier model is only 10 to 20 percent better than an open-weight model that costs 30 times less, so build a small personal eval and pay frontier prices only where you can see the difference.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You pay for one or two $200 a month AI plans and wonder what happens to your workflow when the subscription subsidies end.
You are building an app on a frontier model API and the monthly bill is starting to look unsustainable.
You keep hearing the word evals and have no idea how to make one for non-coding work like writing, design or video editing.
You edit video, make 3D scenes or build documents and want to know why AI does better when it skips the app and works on the files directly.
You are a software engineer or specialist worried that AI is erasing the years of experience you built up.
SKIP IF…
You want a step-by-step tutorial; this is a conversation with ideas and examples, not a screen-share walkthrough.
You want benchmark charts comparing models; the only numbers here are anecdotes from the guest's own tests.
TL;DR
The full version, fast.
The newest frontier model is rarely worth its full premium. In the guest's own design tests, an open-weight DeepSeek model matched GPT output at one thirtieth of the price, and the frontier pick added only 10 to 20 percent. Open-weight models trail by months, then flatten prices by 20 to 30x, so the smart move is a folder of five or six prompts you actually run, tested across models through OpenRouter whenever something new ships. The conversation then widens: almost every work task is secretly a coding task, so AI edits video better through ffmpeg than through an editor's UI. And when AI removes the barrier to a skill, the answer is to aim much bigger, not to defend the old role.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
00:30Guest, founder of Particle and co-founder of MixpanelSuhail Doshi
Chapters
Where the time goes.
00:00 – 01:11
01 · Cold open and intro
Highlights from the conversation, then the setup: a Mixpanel founder now running an open-model company tested the newest Claude model against cheaper options.
01:11 – 03:33
02 · First impressions of the new models
The guest finds the new frontier model refuses a lot of legitimate inference and training work and worries about silent downgrades. The host is using it to edit video and sees non-creators shipping videos with it.
03:33 – 08:04
03 · The GTA benchmark
Why models need ever harder problems. All the pieces for building a real game already exist (3D, CAD, audio, code, networking), so a hard public benchmark like recreating GTA 1 through GTA 6 might be what gets labs to climb.
08:04 – 27:10
04 · AI pricing, subsidies and the business of tokens
Usage keeps exploding with 500-turn agent sessions, 24/7 agents and swarms, while open-weight models flatten prices 20 to 30x. Consumer AI likely ends up ad-funded. Particle post-trained DeepSeek with reinforcement learning to be 42 percent more token efficient and sells tokens on a Stripe-like low-margin, high-volume model.
27:10 – 30:07
05 · When to switch to cheaper open models
Try models side by side on OpenRouter. Companies usually build on a frontier model, look at the bill after launch, and discover they do not need it. Keep your own small eval to find the 30x savings.
30:07 – 33:55
06 · Building your own evals for real work
A folder of five prompts you rerun on every model is enough to start. The guest's design test found DeepSeek matching a frontier GPT model at a thirtieth of the cost. Advanced users can turn a finished session into a scored eval.
33:55 – 42:08
07 · Every work task is a coding task
Video editing, PDFs and Blender scenes are all files and code underneath. Models do better working headless with ffmpeg and Python than driving app UIs, and features like version control are moving into the model itself.
42:08 – 51:12
08 · Adapting your career and following curiosity
Software engineering as a job is giving way to AI engineering. Scale up your ambition, follow odd interests without judging them, and recognize that technology resets the table in favor of newcomers.
51:12 – 1:00:49
09 · Scaling your ambition: work, creativity and robotics
Structure your life so you are allowed to do more, leave jobs that ban AI, and expect video editing and robotics to follow the same 10-month leap coding made from Flappy Bird to a multiplayer shooter.
Atomic Insights
Lines worth screenshotting.
In one design test, DeepSeek V4.1 Flash produced output equivalent to a frontier GPT model at one thirtieth of the cost.
The most expensive frontier model gave only a 10 to 20 percent improvement over a model that cost 30 times less.
Open-weight models usually trail the frontier by months, then cut the price of that capability by 20 to 30x when they land.
A frontier model stays premium for roughly six months before rivals catch up, and the lab has to recover its training cost inside that window.
AI usage has grown from one or two turns per conversation in 2022 to agent sessions running 256 to 500 turns.
A useful personal eval can be five prompts in a folder that you rerun on every new model and judge by eye.
You can turn a finished editing session into an eval by asking a frontier model to score future outputs against what you liked about it.
Most companies die from building something nobody wants, not from lacking a moat.
Tokens map to dollars almost one to one, which makes inference look more like payment processing than software.
AI edits video better when it manipulates files with ffmpeg than when it drives a video editor through an MCP integration.
The UI humans needed to edit video, design or 3D is the worst possible interface for an AI model.
Blender is so tractable for AI because every action in it is exposed through Python.
Features like version control are moving out of software and into the model itself, which will just ask where you want things saved.
Technology resets the table: nobody has ten years of LLM inference experience, so a new graduate starts nearly even.
If your employer bans AI in your role, start looking for a company that is trying to become AI native.
The best way to quiet the fear of AI replacing your skill is to aim at a project that used to need a whole team.
Takeaway
Test the cheap model before paying frontier prices.
WHAT TO LEARN
Frontier models win by a smaller margin than their price implies, so a tiny personal eval, file-level AI workflows and a bigger ambition protect both your budget and your career.
02First impressions of the new models
New frontier models can refuse legitimate technical work, so test a release on your own tasks before switching every session over to it.
Video editing is one of the clearest new strengths of current models, letting people who never edited before publish polished videos.
03The GTA benchmark
Models need increasingly hard public benchmarks; a measurable target gives labs something to climb and turns impressive demos into real products.
The building blocks for full games already exist across code, 3D, audio and networking, so the remaining gap is mostly tuning and data.
04AI pricing, subsidies and the business of tokens
Expect your AI usage to grow sharply as sessions run hundreds of turns and agents start running around the clock.
Frontier labs hold premium pricing for about six months before open-weight models arrive and cut the price of that capability 20 to 30x.
Post-training an existing open model with reinforcement learning can cut token use 42 percent, which lowers cost and speeds up every task.
Build something people actually want first; most startups die from lack of value, not from lack of defensibility.
05When to switch to cheaper open models
Prototype on a frontier model if you like, then check the bill after launch and ask whether the task really needs it.
Use OpenRouter to swap models quickly and compare price against quality on the same prompt.
06Building your own evals for real work
Start an eval with a folder of five or six real prompts and rerun them on every new model before you commit to it.
Once you have a session you liked, ask a frontier model to turn it into a scored eval you can run automatically on new releases.
07Every work task is a coding task
Ask the AI to work on the files directly with tools like ffmpeg or Python instead of driving a video editor or design app.
Treat any creative tool with a scripting layer, like Blender's Python API, as something you can now control by describing what you want.
Organize where AI-generated files land from day one, because headless workflows create the same messes vibe-coded apps did.
08Adapting your career and following curiosity
Retrain your role as AI engineering by learning basics like tokens, transformers and how models are trained; the shift will take decades.
When AI threatens a skill you rely on, pick a project that used to require a team or studio and make that your new target.
Follow strange interests without judging them; the guest learned to program writing bots for a StarCraft clan.
09Scaling your ambition: work, creativity and robotics
Shape your job and finances so you are allowed to scale your ambition, cutting back obligations where you can.
If your company forbids AI in your role, start looking for an employer trying to become AI native.
Expect fields like robotics to open up to people who only know how to describe what they want in plain language.
Glossary
Terms worth knowing.
Frontier model
The newest and most capable model from a top AI lab, usually the most expensive per token and the default most people pick without testing alternatives.
Open-weight model
A model whose trained weights are published so anyone can run, host or modify it, such as DeepSeek, GLM or Kimi. They are typically far cheaper to use than frontier models.
Eval
A repeatable test you run on a model to judge its output. It can be a large scored benchmark or just a handful of your own real prompts that you compare across models.
Post-training
Improving an existing pre-trained model on specific behaviors, often with reinforcement learning, instead of paying the enormous cost of training a model from scratch.
Reinforcement learning (RL)
A training method where the model is rewarded for desired outcomes, such as reaching the right answer with fewer tokens, so it learns to produce more of that behavior.
Token efficiency
How many tokens a model spends to finish a task. A more efficient model reasons less wastefully, which lowers cost and finishes faster.
OpenRouter
A service that gives one API and account access to many models from different providers, making it easy to switch between them and compare price and quality.
Headless
Running software without its visual interface. Here it means letting the AI edit media files directly with code and command-line tools instead of clicking through an app.
ffmpeg
A free command-line program that can cut, combine, convert and render audio and video files. It appears so often in code that AI models are very fluent with it.
RL environment companies
Firms that build training environments and data, often from recorded task trajectories, that frontier labs use to teach models new skills.
Hill climbing
Researchers' term for steadily pushing a model's score up on a specific benchmark through repeated tuning and extra data.
“I noticed that there was almost no difference between GPT-Astra and DeepSeek v4 4.1 Flash. It was the same, but the cost was 30 times lower.”
Concrete, surprising price claim from a credible founder→ Hook for a 'stop overpaying for AI' short↗ Tweet quote
23:12
“Most companies die due to lack of value. They never made anything that anybody wanted.”
Universal founder lesson in one breath→ Founder advice clip↗ Tweet quote
24:22
“Intelligence is as close to dollars as I have ever seen anything. It's like a one-to-one mapping on tokens to dollars.”
Memorable framing of the AI business model→ Business-of-AI explainer↗ Tweet quote
40:43
“The features are in the model now. Not like we are building the features and software.”
Contrarian take on where product lives→ Builder-audience provocation↗ Tweet quote
43:28
“I think software engineering is dead. It's not software engineering. We're doing AI engineering now.”
Strong declarative line already used as the cold open→ Career-shift reel↗ Tweet quote
45:15
“If video editing is going to go away, scale up. What is the thing that you never even thought was accessible to you that you can go do now?”
Actionable reframe of AI anxiety→ Motivation clip for creators↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
I think software engineering is dead. It like completely just chopped everybody off that was working for 20 years, which is scary. I think people are pretty scared of that.
It's always important when the experts in the industry freak out because they know. So Particle, how do you guys make money? The goal of the company is to drive the cost of these tokens to zero.
Today, I'm talking with Suhail Doshi, founder of Particle, which takes open models like DeepSeek and makes them cheaper to run. And he's actually done this before. Suhail started Mixpanel when he was 20, and today it's a billion -dollar company doing over $100 million per year.
So when CloudFable 5 .2 dropped this week, I wanted his take. And I've been using it to edit videos, make 3D films, and build a ton of apps. And when he ran his own tests, Fable was only 10 % to 20 % better than a model that cost...
30 times less. And so in this video, we talked about how to test which model you actually need, why AI edits videos better when you don't use a video editing software, why every job is turning into a coding job, why he thinks software engineering is dead, why he wants AI to rebuild GTA 6, and he also gave me some advice that I can't stop thinking about.
If you're new here, my name is Riley Brown. Let's dive in. Suhail, so...
In the last three days, I guess in the last two days, we got Claude Opus 5, or no, Claude Fable 5 .2. And we also got Gemini 4 Carbon. Have you tried these models?
I'm curious about your thoughts on these models. Yeah. Yeah.
I usually like to, these days. I have stopped being on this trend where like the model comes out and I quickly go to the model picker and I change it or whatever, usually because I already have a lot of work sessions going on. But I did try Fable 5 .2 earlier today and I just started giving it a whole bunch of problems.
And I have not tried Carbon because it's not even clear to me how to use Carbon sometimes because I might be in Codex or whatever. But yeah, I've tried Fable 5 .2 and the best thing that I could say about it is it... really likes to deny the work that we'd like to do with it actually like there's a lot of weird guard rails and i think one of the nervous things that i have about fable 5 .2 sometimes in some of the anthropic models is actually that kind of the frontier sort of work that we're doing which is inference optimization and ai training models and things like that i am often worried that we're getting downgraded because they will just randomly choose like a crappier model and i'm always like slightly paranoid that that's what's happening.
And then we're just getting like worse code. But I did give it a shot before before today and just thought thought I would try it out. And I think I need more time with it.
Probably. Got you. Yeah, I mean, For me, like as a creator, I've been really obsessed with using it for video editing.
I think that's kind of the biggest shift I'm seeing right now. You know, my buddy, Muhammad Zulali, he's an engineer. And he's been wanting to make content for a while.
And he made a video. You probably saw it on the timeline. He reverse engineered his Jeep Cherokee.
Like he got clawed to like control his car. And he... used fable 5 .2 to edit the whole video and it's going to get like over 2 million views on youtube and i'm starting to see that a lot of these people who would have never made videos are just kind of filming themselves throwing it into fable and um so i'm really excited about that that's really good at like yeah it's it's amazing and it's super good at like visual understanding and editing videos and so that's what i've liked the most about fable one thing that you tweeted today that i want to ask you about you said g You said GTA Bench.
What percent of GTA can the model recreate? Start with GTA 1 and hill climb to GTA 6. 100 % accuracy is fully recreate the game that would be indistinguishable from the original version for players.
What did you mean by this tweet? I'm very curious. Yeah, I mean, I've had this long view that these models need increasingly hard problems because I think they continue to sort of astound us.
And I think one of the things that... maybe a lot of us didn't think that they would be able to do maybe maybe around four years ago was editing videos i think and then i think even just two years ago or three years ago we probably thought that maybe the video models would be the ones that edit videos and the same thing is kind of happening it's but it's happening in other industries so one industry that it's happening with is also robotics the roboticists like went into a total existential crisis these last two three months like i spent a bunch of time in robotics last year and i think three different roboticists that you would probably respect, just given their pedigree, all kind of went through like a, oh my god, what does this mean?
The LLMs can do it. So I'm always interested in like, thinking about the things that the LLMs, we wouldn't have expected them to be able to do. And so increasingly, over the last few months, people have been making astounding games.
And right now, I think a lot of the games are maybe a little bit impractical, like they're really good demos. But the interesting thing about great demos in AI is that quickly great demos become like actual products or actual things.
And so I just think that, you know, with GTA 6 coming out, I don't know if you play video games, I definitely preloaded GTA 6. And I was just, you know, pondering that, you know, we, it would be, you know, and I've also been spending a lot of time looking at benchmarks, cyber benchmarks, health benchmarks, all kinds of interesting benchmarks.
And I think that if we could get the models to sort of figure out if we could get gta1 to work right that would be really incredible i think people will be pretty astounded by even just that and there's this other big movement that's going on as well in the last few months which is rea which is reverse maybe it's like reverse everything and anything i'm not it's reverse reverse engineer engineer everything yeah that's right yeah and so that's going on as well and that's also mind -blowing people are making all kinds of weird games where they're instituting their own characters and things like that so it just feels like a really hard eval for video games would be trying to ultimately recreate gta 6 which took how many billions of dollars and how many like a decade i just think that would be fantastic yeah and you mentioned that a lot of these things that we're seeing on twitter are just demos and you're you want like the real game i'm curious what are the limitations of that like what do we need to achieve or what do models need to get better at
to go from demos to the real game in your opinion? Yeah, I mean, it's like we have all the pieces. And that's why I think I made that tweet was because one piece was now we're seeing the models do 3D Blender stuff, CAD, the RL environment companies that generate data to help train frontier models.
People started reaching out to me about being able to do CAD. And so it's kind of like we have all the pieces, you know, we've got audio figured out, we've got 3D models and stuff figured out, we've got coding figured out. for you know networking and multiplayer games it's like all the pieces are there and so often what needs to happen maybe what needs to happen is something even simpler than the technology just purely oh like we need to add more data for this particular thing is actually just a really great benchmark for the frontier labs to climb because often when you create these benchmarks i mean people are a little lazy and so to the degree that someone can make like a really good benchmark for these researchers you know to the degree that there's a number that they can hill climb that's sort of like catnip and so it might actually just be a benchmark because all of the little pieces are there.
And so in order to hill climb the benchmark, it might just be like a lot of tuning and like extra data, but ultimately I think we can get there. So it's maybe just the motivate, like something that motivates them to probably want to go and do it. But yeah, I don't know that there's anything really left.
Like we're on our way. Yeah. I mean, three weeks ago or four weeks ago, I was using Opus 5 .5 to make a game live on stream.
And then in between, like it was, it was multiplayer. So we just had just, it was convex real -time database with just hosted on Vercel. And we were playing a game and we would play for 20 minutes.
And then I would just talk to the chat. And then we were like, what do you want to change about the game? And everyone would comment in the chat.
And I was just copying and pasting them and just ran it as one prompt with Opus. And then the game would update. And then we would play the next version of the game live on stream, which.
That's amazing. That's awesome. Which I think is just like the future of gaming.
But I want to get off gaming a little bit because I think one thing that I can't stop thinking about, and I think about this all the time, is the off -ramp from all the subsidies we're getting by Anthropic and OpenAI. You know, I've been using, I have the $200 per month plan on Clod. I actually have two $200 per month plans.
It feels for me, because I'm like doing 80 % creatives work, 20 % coding stuff. To me, it feels like I get unlimited tokens. However, when I was at my previous startup, I accidentally used API pricing for one month and accidentally just spent $12 ,000.
And so what I'm concerned about is like this off -ramp period from when we go from all these subsidies to not having... the subsidies where, you know, we pay closer to the API pricing. How do you think this plays out?
Like, how do you think Anthropic and OpenAI are going to kind of, are they just going to slowly reduce the amount of like credits we get on the paid plans? Or how do you think that's going to work out? Yeah, so it's very interesting.
So I think there's kind of like multiple angles to it. The first angle would be just like our own usage. I actually think our own usage.
is going to dramatically grow. And I'll give you like an example of how, I mean, if you just sort of look at the lineage since 2022, ChatGPT came out December, November, November 2022. We started to use it.
And most of ChatGPT was like probably like one or two turns, right, with it. You'd say it's, you'd ask something, it responds, maybe you're done. Maybe you ask one more question, you're done.
And then we started getting things like tool use and search. Then it became a little bit more multi -turn, maybe three, four turns, but then you'd stop. And if we just fast forward a little bit, so I don't have to go through the entire lineage.
I mean, we're looking at like, I see trajectories where, you know, if you look at like a trace of someone and what they're doing, I mean, people are taking like 256 turns, 300 turns, 500 turns. These are becoming really long horizon tasks. So that's one element is that we're just doing way more with AI.
And then the second thing that I think is probably going to come are these sort of really frontier things that people are not really doing yet. One frontier thing that I think people are not really doing is they're not running like an agent that's 24 -7. For example, one very obvious situation where we should be doing that is in cyber.
We're not doing that. And then another situation that I think is a little bit too frontier because the AI companies haven't really released models that enable us to do this is we don't really have a great easy way to do agent swarms yet where we like literally spawn a thousand agents to go do something. I mean, a lot of that requires the model to be somewhat post -trained.
So I think on one end of the spectrum, we have not yet quite realized the quantity of tokens we might use actually, given if the models sort of improve and things get better. And then the other aspect is just pricing, right? And I think that pricing has shown that about every year, the prices have gone down relative to the same unit of intelligence.
The reason why... Prices have actually kind of gone up in some years, but relative to intelligence, it has gone down. And so some prices have gone down, but a lot of the frontier labs are actually pushing higher prices.
And so a lot of users, as a result of that, just pick the best model, but they're actually paying more and they don't even know, is there a cheaper model? It's actually too complicated. There's like five different models and then four different versions of reasoning.
And it's just too much for the average person, I think. So I think that's another element. And then the third element I think is that I'd find most exciting is actually all the open rate models that have come out.
And so one thing that people ought to think about is that the frontier labs have earnings before training, which is that they make money and then there's the margin they make from inference. And then there's the margin that they make after all the compute that they had to use for training. And that cost of that is actually the period of time that they got to train that model and the success of the premium dollars that they got for the period of time that the model stayed premium, right?
Fable 5 .2 will not be a premium model in probably like six months, right? So they have to eat that cost. They have to stay premium for some amount of time until OpenAI or DeepSeek or whatever, ZAI comes out with their models.
And so the other thing that I think is happening that's really exciting and I think really great for the world. is the open weight models are coming out and then just like flattening the field every time. And they are often a bit behind, but they flatten the field so much that there's this 20, 30X price reduction.
And so I think a lot of the open weight models are eating up the frontier models. You know, whatever was frontier last year is no longer frontier task and these open weight models just eat everything up. And so if you look on OpenRouter, for example, it's growing so fast like i think deep seek is you count both versions of deep seek that are like roughly the same model it's 60 70 trillion tokens a week growing exponentially actually got you there's so many directions i want to go with this um what i want to talk about first is you were talking about people running agents 24 7.
And so obviously, I've talked to a lot of developers who are running agents a lot of the time, right? And they're doing a lot of these cybersecurity workflows where they have them run basically all the time. They're checking things.
Anytime someone posts in Slack or Linear, it triggers off some workflow, which is interesting. I think a lot of people are talking about this. What I'm curious is how do you see this playing out for businesses and then just end users and consumers?
I think we're only just now. kind of in the early stages of, you know, consumers being able to do this with tools like the Hermes is the kind of the open one. And then you have like GPT dots and Grokbot, et cetera, which gives you a persistent computer so that you can actually run more work tasks 24 seven.
How do you see, like, do you think in a year, a large amount of people will be running agents 24 seven? Do you think they're going to be setting them up or do you think there'll be set up for them? Like, how do you think?
that's going to work for average business people i think on the consumer end i think it's going to go the way of things like search like could you imagine paying for search you know no way and and so i think the way that it goes on consumer is that actually people don't really want to pay 20 a month on the consumer end 20 a month is a lot of money i mean if you've ever run a consumer company and you charge people you know 15 a month i mean it is like bloody murder if you charge them a month and they didn't expect that charge So I don't really think the average consumer really wants to pay for AI at all.
And I think that is going to get totally subsidized. Someone is going to find the right business model. Maybe it's you make money off the e -commerce.
Maybe it's money off of ads. Likely it's, you know, generally it's ads. And so I think that that case, it's just a function of like, what is the capability of model that you need?
And once we reach this kind of like threshold or plateau rather of capability, right? Do you really need a frontier model? for non -frontier tasks probably not and so a large portion of consumer tests i think are going to be non -frontier and you know it's not like they're using the models to go and like research cancer or anything right so so i think a lot of that that cost is going to hit a a bottom there's going to be a point where the the price decreases so quickly uh over the course of the year or two that there'll be this moment of a threshold where like it will unlock a ton of consumer categories that were previously like i don't know how to make money off of this because none of these consumers would pay us up So I think that's probably the realm of the way I suspect consumer will go.
And we'll see what happens. It'll be really interesting. I think most of the token dollars will flow towards enterprise.
And we will increasingly find the only way for companies to have alpha is that they have to come up with really hard things to do to compete against each other. And they require time and energy and capital and focus and those kinds of things. And so I think a lot of the tokens will end up with an enterprise.
I don't know if I answered your question, but I just thought I'd pause and see what direction you want to go. Yeah, it's very interesting to see all the different players on the consumer side. I know you have Instinct, and they're raising literally billions of dollars.
And I guess their business model is taking a percentage of the commerce that happens on the platform. I don't know if that's true. I heard that.
And then I think OpenAI's strategy is to... is to do ads, which makes a lot of sense. I'm curious.
So what I want to talk about is, is Particle. So you are trying to make models more efficient. You guys, so Particle has trained, you guys trained your own model, correct?
Yeah, we trained our own model. Yep. And so you trained your own model and did you use one of the open models as, as like the base for the model or you want to just talk about that process?
Cool. Yeah, so one of the things that we've kind of learned is that things like pre -training and pre -training a model is very expensive. That's the kind of thing where you need to get a lot of compute and you need to raise tens of billions of dollars at this point to really go do that.
And so there's some labs that are trying to make that cheaper, but ostensibly to compete at that level, it's a lot of money. So what else is there? Well, the other thing that we figured out was that we can do what's called post -training, which is that you take an existing model that's already been sort of made like DeepSeq or GLM or...
inkling there are a lot of different open weight models out there and you basically just like tell that make that model like better at other kinds of tasks and so it just turns out what we've learned is that the insight that we kind of learned is that post -training is actually like this very long wide surface area and it used to be that when you train these models you know you do a little bit of post -training and you're done um it was kind of like the final touch if you will But that's that to us, that's like no longer the case.
And so what we did with particle one is our first model is we did something that we noticed in the market, which was that a lot of these models are very token inefficient. They just sort of overthink and they over reason and spit out a lot of tokens. And and so, you know, we we were wondering, like, could we actually make it more efficient?
Could you know, people have a cost sensitivity to using all of these tokens, whether it's personally or at work. to what degree can we lower costs in a different vector? A lot of the inference companies that we may compete with or are rivals of ours, they mostly focus on writing GPU kernels or just making the price lower.
But we thought, what if we just make the model more token efficient? And so Particle 1 basically made...
made it 42 % more token efficient than DeepSeek. And we actually saw people that were kind of complaining that DeepSeek uses a lot of tokens. And the other factor is that by being more token efficient, you actually complete the task roughly that much faster, which we think is pretty cool.
I know that I don't love sitting in front of Codex just like waiting for these things. That is the one thing I hate the most is how often I'm just like spread between like four different things at once. And then I just end up waiting for it to finish.
I'm curious how can you give us a breakdown on like how you make a model more efficient? Like what does that process look like without spoiling any secrets? Right.
Yeah. Yeah. I mean, it's we just use reinforcement learning.
You know, that's the way we do it. We're not doing anything where we're modifying weights or anything like that. It's just really good data and really good empirical kind of like science and evidence and reinforcement learning and sort of picking the right shape of your algorithm or reward rather when you're doing RL in order to like get the model in the right zone.
And so one of the things that I think that we, one of the things that I think is sort of boring, but actually really important is just like really doing a good job. finding the right like eval that you would use and so we spent a lot of time actually looking for like the right kinds of data and and so i personally even i personally spent days just like looking for the right kind of data that would adversarially like show us bad token efficiency and then we sort of adjust adjust the model and use reinforcement learning to to do better on those data sets and since launch like we found more like some customers have had really great uh have kind of like independently shown showed us that we actually do do it and then some customers are like hey like i don't quite see the gain in this particular category so now we've been kind of like adding adding up all of the feedback so that we could go do round two so you started with deep seek right deep seek yeah which 4 .1 flash yeah deep seek 4 .1 flash 10 17 for those that care forgot got you got you i'm curious like um where does deep seek fall in i guess
Like, can you give us a breakdown on like the different open models and like the different players there? Like if I were to go test, maybe I wanted to use Hermes agent or one of the other platforms that lets you actually use the open models. What models should we start with if there was like a list of four?
Well, I could even cut down the list of two, but I think the two great models, I think that from my favorite two open wave models. are actually the ones that come from DeepSeek and GLM, which is a company called VAI that's behind them. They have both Pro models and Flash models.
Flash models tend to be the cheaper ones. Users really like those two models. I think they're just very high quality, not overly benchmarked, that sort of thing.
Whereas I think a lot of people do like Kimmy. It's a little bit more benchmarked for whatever reason. And maybe keep it to those three.
I think those three are pretty good. A lot of people like the... like the 3 trillion parameter Kimi model for some of the more frontier workloads.
Very curious. So Particle, how do you guys make money? I'm like very curious about, I'm very curious about the model.
Like I've just been, you know, I was hit with a wave of, I guess. not cynicism about, you know, creating software and creating wrappers just because it feels like the labs are creating everything. And I've talked to a lot of SaaS founders who are like really struggling, especially the ones who don't really deeply understand marketing.
It's just become like a marketing game. And I've been just trying to kind of understand, I've been really trying to understand the business model of the rest of the, like going deeper down the stack. And I think this is a completely a new business model that i'm not familiar with so i'm just really curious on like how you guys make money and like how you plan to make money in the future yeah you know i have a very simple thesis on making companies these days i try to like actually try not to overthink it too much i think there's a lot of overthinking that you can do especially if you're doing it like a second second time i call like second time founder syndrome and i think one of the things that that is just hard to do when you start a company is the hardest thing actually is not there are very few companies that die because of lack of moat.
Most companies die due to lack of value. They never made anything that anybody wanted. And they were so focused on defensibility that they like invented a product that nobody actually wanted.
And so I think to answer your question, we make money off of tokens. It's not that complicated. The business model is sort of simple.
We make money off of tokens. Now we have this funny aspect to the company where the goal of the company is to drive the cost of these tokens to zero, which belies this question of like, well, then how might you make money in the future? But what I find...
That's why you asked. That's why I'm asking the question. Because I saw your tweet that said that.
I was like, okay, if you're trying to drive the cost down to zero, it's like... what's the business model? Yeah, yeah, yeah, totally.
Well, you know, the thing is, is that I think it's gonna take time to drive it to zero. I think that's one element of it, right? I think the second element, so for a while, it'll be okay.
I think the other element, and so we don't mind making kind of like not that much money. There's also an analog to this business. Never has there been a business, I think, since payment processing companies, where an increase of a unit of your thing.
converts directly to dollars. Like intelligence is as close to dollars as I have ever seen anything. It's like a one -to -one mapping on tokens to dollars.
It is, it feels like payments. And so I often liken the company quite a bit to like Stripe. Shout out to Patrick.
He's also an investor of ours, but he's, but Stripe, you know, what Stripe, what Stripe, what margin does Stripe make? Less than 2 .3%, right? in reality right their big enterprise customers probably don't pay 2 .3 percent and so stripe is definition of a commodity business um with payment processing right and intelligence is much better margins than that at the moment and and so one one of the things i think that is is maybe worth worth mentioning is that intelligence though it works to have a low margin business but only in the situation where the volume is like ever expanding like nearly unbounded like payments payments is just like infinite right or whatever the world i mean it's finite to a degree but it's the pie is constantly growing and intelligence is also uh has a lot of those same feelings it's just this unbounded thing that keeps growing exponentially year on year right now and so it kind of doesn't matter to have low margins and so that that's one element and i think the other element is that let's say we were to drive the cost of tokens down to zero
Well, if you achieve that, then our company will have amassed as many developers as we have. I mean, we have a ton of developers using our products. And so I think one reason it can be hard to sort of say, well, we'll do this is because it's really hard to predict in five years or 10 years this thing, because it's super hard to predict that.
But there will always be something that is next. And so I suspect that what we'll actually have is just. huge distribution of developers where we're just trying to help them, again, build better applications with intelligence, where intelligence originally was the business model and now it's going to be something else.
And I just can't perfectly predict that. But I don't think that I need to because I suspect there's always going to be something that is using these AI models to help build better applications and things like that. It seems like CPUs ought to be zero, but they're not.
They're still not. That's a really good point. You don't have to predict it now.
you can just help people make or use intelligence at a cheaper level until something will present itself. Even 18 months ago, it was hard to predict how people in their job would be using AI on a day -to -day basis. But in hindsight, it seems obvious, I guess.
Yeah, GPUs were just like that. They were used for games, and then suddenly they found a use case in AI. And there's no way that the folks at NVIDIA predicted that was coming.
the first two years of their company, let alone the first five months, right? And so I just think that the first order bit of making a company is actually just do something that's valuable to people. That's hard enough.
Right, 100%. And okay, so for a lot of my tasks, I'm using frontier models. How can I, as a small business owner, start to transition from using frontier models to using cheaper models.
And I guess another way to ask this question is like some tasks somewhat require the frontier models or should be used. I guess if I'm running everything on like, like basically everything that I do is set up through some sort of automation using either cloud or codecs. And I don't know which platform that I should be switching to, to use these, these open models.
And so. Even if I wanted to use Particle or one of these open models for a lot of my work tasks, what tools do you recommend using at a more practical level? Yeah, I think OpenRouter is a great place to try stuff and experiment with it.
It's undoubtedly the best. Obviously, we say, oh, people should come to our site. We have these open -weight models.
But if I were to be totally honest, I think that OpenRouter is a great place where just quickly flip between them and give it a shot. I think one big part is that because AI is so nascent and... and sort of new to the world.
In a sea of a lot of options, you trust the brand. You pick a brand you trust, right? And I think that that's definitely happening, whether it's what harness you use, what models you use, what ADE you're using, right?
There's so much. And so you're just like, okay, what is everyone else using? What's roughly the right direction?
And I think that's a reasonable thing. And then I think there's a secondary bit, which is why we're not chasing after like codecs and we're not. We're not trying to build a business in that lane because we think there is actually a lot of confusion there.
We tend to focus our company on serving developers who are making applications using AI, where they have very specific requirements around like, this is what I want my app to do. And it has to have this performance. And once they lock in that, they might use whatever, they might use the frontier model.
But after they get there and they ship that in the product, they look at their billing. And they go, oh, my God, this is unsustainable. They ask the question, do we really need a frontier model?
And the answer is often no. And that's where we think that we can kind of get involved. And also, they're able to use OpenRouter to switch between these models and try to get the right.
price performance ratio. So I think on a personal level, like even I use the frontier models, and we just haven't built the right UX around knowing really when to switch. And we almost don't trust it.
We're like, Oh, no, if it's using Luna, like maybe I'm not getting as much mileage out of if I was using Astra extra high, right? If you're in if you're one of these, if you're as nerdy as I am around these sorts of things, but I think when it comes to using it in your application or running something day to day, I think that I think that just you got to just like try them and look at look out for evals.
And you should always have your own eval, even if it's five things, six things that just really measure that, because you could likely get a 30x reduction in price for the same thing. And I think a lot of people just don't realize that the open weight models are really good at things like that. Yeah.
And where can we find or create our own evals? I hear the word evals a lot, honestly. Yeah, I hear the word evals a lot.
And it makes sense for coding. But a lot of it feels imaginary. Like maybe I just haven't taken the time to actually learn evals.
Maybe it's intimidating. I don't know why. But like, how can we create evals for non -coding stuff?
I mean, that's been my main focus right now. And a lot of my audience are people in businesses, you know, like maybe they're, you know, a marketing executive. For these types of tasks.
I feel like people are kind of defaulting to the frontier models just because they trust them more for whatever reason. It's hard to kind of build trust to start using the open models. So I'm curious on how we could actually set up evals to measure how good a model is at something like video editing or something like script writing or general administrative tasks.
Like, what are your thoughts on that? Yeah, I think, I mean, I think this is different for every person. So on a person level basis, you know, there's Simon, I think his name is...
I forget what is Williams maybe. And he always has this. Well, Simon Wilson.
And so he has this, he has his Pelican eval, which he likes to do. And he probably has other ones, but I think that, I think that that's the right shape of things that you sort of take something that you find yourself doing quite often in your case, that's like a video. And so you just have a folder and you just call it evals and you just copy and paste the prompt from one model to the next.
And you just go like, is it better or not? And one thing, I mean, I did this for design for a little while and. The funny thing is that I noticed that there was almost no difference between GPT -Astra and DeepSeek v4 4 .1 Flash.
It was the same, but the cost was 30 times lower. And the result of the design was equivalent. And maybe I got a marginal 10 -20 % improvement using Fable.
So I thought that was really interesting. And I think just having a small little folder... of your things and small business of like okay i'm going to just like run the models on these five things and that should give me a barometer reading of like this model relative to the things that i like and then i think for the more sophisticated people you know that you could probably use the models to make your evals like use the frontier model to help you you know sort of like you do your trait you know you have your like trajectory where you edited a video and you went back and forth and you could probably just like and then you get to like a happy place with your video right and so you have your result and then you could probably just download that trace if you're a little bit more savvy and Ask Fable 5 .2, hey, would you turn this into like an eval where, you know, there's a score behind this.
And these are the kinds of things that I liked about my video. And can you make it so that it can score those pieces? Something like that could really help also.
And that might help you start automating an eval where you can just run it when new models come out and then just do like performance to price. Yeah, I guess no one's really made like a super user friendly eval platform of sorts. I wonder if that would be.
That could be kind of interesting because, yeah, I think a lot more people are starting. Go ahead. Well, I was going to say that the industry that we're thinking about here with Traces is actually this big budding industry called RL Data Environment Companies, which I'm sure you've maybe come across here and there.
And so it's almost like a little secret for you to not like it's almost like they don't want you to know how to do it or at least how to construct these evals as well. So that might be in part why it doesn't exist. Yeah, I mean, it seems.
I would like for it to be more accessible because you're basically just taking a prompt and you're just evaluating the output and you're measuring across different models. But I guess I've actually tried to set it up and it always ends up being unclear. It takes a lot of work to evaluate whether something is good or not.
It's just there's a lot more tedious stuff than you think if you're going to set up evals. And I think that's where I kind of get stuck. You should make a skill.
I bet someone could make a skill that gets everyone 80 % of the way there. Take this trajectory and turn it into an email. I saw you tweeted the other day.
You were talking about how you're interested in working with people who are kind of building co -work -like tasks or co -work style. I'm curious on what use cases are you seeing these companies use your model for? And where are open models good at for co -work style tasks?
It's always hard to tell the line between what is a coding task and what is a work task. They almost feel like they're the same. Good coding models end up being really good at work tasks.
I'm curious what you think there. I agree with you. I mean, I think if you go look, if you're an AI researcher and you're trying to make your models better, you will go look at benchmarks.
And so I imagine most people watching this maybe aren't AI researchers, but... roughly what happens is they go look at the benchmarks and they really go read through them and and so what you sort of realize when when you do this is that you realize exactly what you're saying which is that most tasks that seem like not coding tasks are in fact really just using large amounts of code and and so because the models got good at code they really latch on to that and and so i think that you're exactly right in terms of like why we're interested in that go ahead no no i i was gonna say like that is what i realized about video editing um is At the end of the day, video editing is just manipulating these files.
And I think people have gotten habituated to like think of video editing as not coding because it's all timeline based. Like video editors just see it as timelines. They don't see it as every single thing as a code.
Even a sound file is just a file on your computer and the models can actually directly manipulate those files. And what I've found is that. These models are better at video editing when it's fully headless and the apps themselves are bloat.
When it has to use the MCP for DaVinci Resolve, a video editing platform, it ends up doing a worse job than when you just say like, do it headless, edit the files directly because this is what I've just found. These models, Codex especially, is really good at editing videos fully headless. at the end of the day it's all it's all a coding task and it just uses like ffmpeg to like render the video and it puts the sound file and yeah that it's super interesting i think the reason behind that is really just that the ai models have probably been more trained on ffmpeg than like the computer use interaction or the mcp sort of like api that's connecting to davinci resolve like there's less of an integration during training but like there's so much code that uses ffmpeg so And then FFmpeg is just so powerful as it is effectively an API that it's like this unbounded thing to the model to be able to use as a tool.
And so I suspect like to some degree that's happening. And so what really what's funny is like we as the humans needed this UI, but then the UI, obviously, I mean, we've all used the UIs. We're like, oh, I wish there was this feature.
Why does the knob not go down all the way? And then the irony is like that's the worst possible interface for an AI model because the AI model doesn't care about a UI. It just actually wants to be exposed to maybe everything.
And so I think that's like maybe really interesting because I think what you're saying definitely applies to like every category across the board. Yeah, like I've noticed it with PDFs too. Like everyone like, oh, when you want to create a PDF, a lot of people in work, they would go to Canva.
And now if you try to go to Canva, they have this integration with Codex or Claude. And then it just, it doesn't work because now it's working and it's putting things like, but if you ask it to just make the PDF directly, it's mind blowing. And then additionally, Blender, which has been my biggest obsession, I posted a video, it got like a million views with Fable 5 .2, where it creates these 3D environments in 4K, like it is as high quality as you can.
It uses, I'm learning about it now. this might not be fully correct. It uses, when it renders, it does it in cycles.
So like the lighting is super realistic and it'll like render the video. And I use something called RunPod and RunPod rents GPUs because it would take like 10 hours to render on my MacBook. It just rents it.
And then for, and then it just does all of this. And then apparently Blender, the reason it's so good at Blender is because every single thing that you can do on Blender is mapped to Python. And so it's all just Python code.
So I realized that every single thing is just coding. So yeah, I think you can hear it in my voice. I'm very excited about these creative things.
And I just realized it's all coding and it's so fun to use. And then the interaction now has changed too, right? Your UI is not like this blender thing you have to go do, you know, 100 hours of YouTube tutorials on.
It's just... it's just your voice and sort of like learning like learning like the high level bits of like what you want to do and and then getting it to like construct you i mean you're basically on track to making your own mini pixar it sounds like i mean i yeah one of the problems that i think a lot of people are struggling with um is like version control now because like adobe and all these tools kind of handled it and now people are editing editing videos and they're getting clawed to do it to like edit their video, their video will pop up on their screen and then like the window will go away.
And then they're like, I have no idea where this is on my computer, like where any of the files are, despite the fact that like they plugged in their camera and it pulled all the photos out of their camera and that's what they're using. They don't know where it's happening or anything. So there's a lot of problems that come from doing things headless.
A lot of these same problems are the same problems that people ran into when they were vibe coding, right? When people vibe code. They don't really know what they're doing.
It's not like they are checking to see if the backend works properly. They just have the front end of the app that they can look at. That's why front end stuff were so popular amongst the non -technical crowds because you could evaluate how good it did, but it leads to a lot of messes.
So I think a lot of cleaning up, I think. We need to figure out how to like reorganize all of this. And this goes for every work task too, is like, how do you create good version control?
And I think Claude Anthropics trying to do it, they're releasing a bunch of different types of artifacts and they're going to be like more team centric and they might bake that in. The old software world, you all of us would say, you know what, let's say called two years ago, we'd all be like, oh, you know what we need to do?
We need to build a feature for version control. And I think that's like a boomer attitude. I think the new attitude is like, oh, this doesn't happen.
oh well let's just we'll just train the model to be better at better at learning to do this for you like it should just like tell you would you like to version control this or in fact we already version controlled it we've already put it in this like area or you connect dropbox or whatever it is that you do and it's like would you like to version control this in dropbox you know i noticed that dropbox is running on your computer and it's crazy but i actually think you know there's the adage of like the best ui is no ui and i just actually think that the models are just gonna get so smart They're just going to know that you have these various file systems, iCloud, whatever, and it will just ask you, would you like to do this?
And by the way, not only will it say, would you like, it'll say, well, we already saved five versions. So how about we just like save only this one, but is it okay to delete this other one that was like from yesterday that you're not really, that was like V1? I just think the models are going to get super good at this.
The features are in the model now. Not like we are building the features and software. And I think people haven't quite reached this mentality that actually product is like more in the model now than it was less software now.
I think people are pretty scared of that. They're reluctant to believe that because, you know, a lot of people get their sense of meaning in building features. You know, they want to be the person to be like, oh, we should do that.
So did file clerks. But then we got IT. Yeah, true, true.
That's true. We're not putting little papers in files and then putting them in cabinets. What happened was we got computers and this huge industry spawned around IT that was even bigger than being file clerks.
And the problem with that fear is that people feel this sense of fear because they can predict the demise of their role. But what they cannot predict is the rise of something new. And those people that were file clerks definitely probably didn't end up in IT either.
But a bunch of people that were kind of nerdy were like, oh, I could do IT. That's kind of fun. Okay, so what's the advice here then?
I know we can't predict it, but if you were telling someone some advice to internalize this, how do you approach this new world? Can we try to envision what's coming next? How do you take that mentality without going insane?
Well, I suspect tons of people that watch your YouTube channel are these deeply curious people, right? I mean, they're... They're watching you because you're always talking about this cutting edge stuff.
Like those are everyone watching your videos. You don't even have to worry about because they're so curious and they're so open minded and they're so interested that they will find their way merely because they're curious about something really cool and kind of like edgy and technical. Right.
I think it's I think that those people are going to be great and they're going to be fine. And even without knowing what's going to happen. And so I just think the answer is like, just like.
keep evolving like i think for example all the software engineers that had their grieving period from the month of november to hopefully now hopefully people are done grieving they're moving on on on some level i certainly had a little bit of a grieving period from january february march i was like oh no like what does it mean i used to write code in seventh grade now what does it mean and i think that i you know i think it's just like reformatting your view like instead of thinking about software engineering i think software engineering is dead It's not software engineering.
We're doing AI engineering now. That's the new role. And so like, yeah, you might have to re -skill and learn a new thing.
You might have to learn some, a little bit of AI research. You might have to learn what a transformer is. You might have to learn how to train a language model.
What are tokens? What's a KV cache? Yes, these are all new things that you didn't know that you might be a little bit behind on, but it doesn't mean you can't catch up because this is a 20 year, this is a two decade, three decade, like kind of continuum.
for everybody. So there's plenty of time. This is actually the earliest.
You're still early. It's good to hear you say that. And I think that I agree with you.
A lot of my audience is super curious and interested. At the same time, they are worried. And I think, and it is hard because I get worried.
I am very much like... I'm in a good position. You know, I make content on YouTube.
I have a large following across four platforms. And every once in a while I wake up and I'm like, is all content creation going to be replaced by AI? Like, what am I going to even do?
And I would say I'm more optimistic and more focused than most people. So I just, I know it's very easy to go down the kind of the dark mentality, but I think you're right. I think it's going to be a long, so I guess to kind of like dig into that just a little bit, I guess tactically.
Right. Yeah. What are those things that people should learn about?
And I know this is it's really broad, but you brought up a few topics. Like, should people learn how to train models? What are the things that people should be learning?
If you're a software engineer, like how do you start making the pivot into whatever AI engineering? Yeah, I think that I mean, OK, one thing one thing I want to mention is that one way to quell your fear around, you know, is this thing going to go away because AI is going to just do it all? I think one way to quiet that.
And one way that I personally used to quiet that is that I tell myself to scale up my ambition. If video editing is going to go away, scale up. Like, what is the thing that you never even thought was accessible to you that you can go do now?
Like, you probably think, I could never make a movie just by myself, right? I could never get a movie in a theater by myself. I could never get to Sundance or whatever it is.
massively scale up you know i know i could never have made a children's tv show like bluey i would scale up your ambition because it means that a lot of the things that actually took a lot of people previously or a lot of money now don't and now a lot of people can do way more and i think this harkens back to my gta6 benchmark which is that it's just this idea that like i i never thought making grand theft auto would ever be accessible to me and now i think that my son it will be accessible to my five -year -old you know in maybe a few years or six years it'll be like all of his friends in class are like making gta3 for fun you know and so i think that and he'll have an element he'll have to scale up his ambition he has to go bigger than that then yeah but he won't even care that he'll his mind is like is like that's just default mode right and so that's why i think it's a little bit we have to realize we have a little bit of a boomer mentality because my five -year -old doesn't even care right and and so that's what some of us are going up against
is my five -year -old, but in five years. And I, and so I think that's one thing. I think Paul Graham has a great essay on kind of touching on what you're talking about.
He has this wonderful essay that came out in 2023, which was how to do great work. Just Google it. And, and it basically, you know, I don't want to spoil it for you because it's such a good read, but, but I think, I think a big part of it is just like sort of following these like strange paths, these unusual paths of your interests and not like not judging yourself for having them.
right and then just seeing what happens and then you know when you're young it's really easy to adapt you know it's like how did i find programming i did find programming by like going to the library or taking a class i found programming because i was playing starcraft online and then there were clans and then people made bots and then you had to have a bot if you were part of this clan and i was like how do you make bots and then i started learning about visual basics 6 to bot to find bots and you know fast forward now and now we're doing ai companies but i think that i think just following these little strange directions, particularly when you're young is healthy.
I agree. And I think all of my most effective times, I think in the last like four or five years has been through just like a deep obsession about random things like i got super deep into the mid journey discord in uh 2022 and that's what set me off into saying to ai that's that's what got me into the gpt3 discord before chat gbt came out and then i was just obsessed from there and i think obviously some of that is luck you know it's just like ai's like the coolest technology ever but i think I think it's, it's easier said than done to, to be like obsessive and curious and follow your curiosities just because people have like demands out of life.
Like they, or, or they have responsibilities, you know, they need to feed their family. And so it's just interesting to me to see that a lot of young people hate AI and maybe it's a lot of it as propaganda, but like, I mean, it's just insane. Like a lot of when people give speeches and talk about AI at their commencements, they're getting booed.
When, Kind of like the system feels like it's against them, but like AI is kind of this great equalizer where like it's speeding up. It's forcing people to have to adapt to change faster.
And theoretically, the younger people who, you know, have less responsibilities, they don't have all these deep -rooted habits in an industry for a long period of time, they should be the best suited to kind of win this game. They just... for whatever reason, culturally, maybe the AI industry failed.
Yeah, I think there's a different way that young people can look at AI. And I think that one of my favorite things about technology is that it does a reset. Like, whatever the chairs around the table, whoever the kings around the table were on one generation of technology, completely reshuffled, right?
Like, Sam and Dario were not at the table 10 years ago, right? Now they are. And I just think that one of the nice things, but this actually trickles down in a lot of other ways.
Like one nice, one beautiful thing about technology is that it resets things. And so one way to like sort of the other version, the alternative universe, right? If we don't have this, is that you like go to school, you graduate.
And then what you have to do is you have to tirelessly work for 30 years, climbing a ladder and you'll never be as good. As the other person, because they just put in raw more hours, it's like 10 years times 40 hours a week or however much they work, right? The average person, you just have to catch up.
But the nice thing about technology, and especially right now in AI, is that it like completely just chopped everybody off that was working for 20 years, which is scary. I like totally admit that. But for young people, I mean, it presents a tremendous opportunity because now that person, how many years of knowing AI do they really have?
I mean, if I'm looking for people to do inference optimization, right? You know, it's not like I can go to recruiters and, you know, there's that like meme on X where it's like, oh, I'm looking for someone that's, that has like 10 years of inference LLM experience. I mean, there's just nobody.
And I remember that was like a meme when the iPhone came out and the app store came out. It was like, how many people had more than a years of experience making apps on the app store? Like very little at that moment, right?
I think it presents tremendous opportunity for young people. I agree with you that it's like a flattening and equalization, that sort of thing. Yeah, 100%.
And yeah, it's mind blowing. And I keep thinking in my head about what you said, like scale your ambition. And I think, yeah, you definitely need to put yourself in a position where you can scale your ambition.
I know you probably spend a lot of time with founders. So you're a founder, you're selling to founders. And so these are people who have really set themselves up.
to be able to scale their ambition like you have to be a person to scale their ambition um and i think there's a lot of people in jobs where they're not allowed to scale their ambition at like a structural level almost like um you know their job you know i was talking to someone where in new york where their full -time job was like they would do color grading they're not an editor they just do color grading and then this a little bit of the sound design and that's like their whole role they get like handed footage and then they hand it back and like they're not allowed to do anything else they can't edit the video in full and so the correct answer, if they had unlimited freedom, is to make the whole movie.
Like, figure out how you alone can, like, get the resources to make the entire movie. Because you, with Claude Code, could figure out a way to do all the planning, go out and get shots, like, and then you could organize all this footage and go through it ten times faster, but you're just not allowed to do it. And so, whenever I give people personal advice, it's like, you have to, like, structure your life so that you're allowed to scale your ambition, if that makes sense.
And it's always easier said than done. I don't know if any of this resonates with you. Of course it is.
It's obviously hard if you have a family and a mortgage and you've got to pay the car bill and that sort of thing. But I do want to say, when I was 20, I had nothing. while i had less responsibility like i didn't have a child i also didn't make any money like i made practically nothing so you know it wasn't like i had i had a lot of opportunity either and and i didn't even have a car so i didn't have a bill for a car but i didn't have a car and so i think i think just there's some realities where like you know to the degree that you can put yourself in a position one way to do that is to also like scale back right scale back the things that cause stress and responsibility i think there's this beautiful thing when you're young where you have like almost no baggage and responsibility.
And it obviously gets harder when you get older. But I think that yes, I think I think that you should you should scale your ambition in that way. And I think I'll give you I'll give an element.
I know this is harder. It's harder. It's easier to say this than it than it is to do it.
But I think if it'd be like if people were in engineering jobs in Silicon Valley, and the company were like, you're not allowed to use AI, you have to write this code. And what would anybody in Silicon Valley say to that person? They would say, you should quit that job.
That is the dumbest job imaginable. You should quit that job. That is not the future.
And so I think if you're a color grading person and you're not allowed to use AI to build more and get more out of the company, I'm not recommending you walk into the company and quit your job immediately, but you should definitely start looking. You should be looking for these companies that are trying to find their way to be more AI native because what will likely happen is if they have the right mindset.
you're going to be doing way more than color grading. And they're going to be doing, your teammates are going to do more than whatever they were doing. And you're going to actually be, as a group, you're going to be doing way more.
The team itself will be scaling its ambition to doing, you know, 10 videos, you know, a week instead of five, you know, instead of one, or trying to make a movie or whatever it is that they're doing. Scale the quality. Yeah, yeah.
Like you can always scale the quality. Yeah. And that's the thing that I'm trying to do.
Like I figured out the quantity game. And it's very funny that we're talking about this now because uh, starting next year, I'm working with filmmakers in New York and I'm going to turn my Instagram into mini short films.
And I'm like, you know what? I can do this. Like I need, you know, I was, I, I was got a little nervous because so many content creators are making content about AI on Instagram.
It's like harder to stand out. So I'm like, okay, what can I do? And I'm like, you know what?
Like, let's make short films. Like it's always something that I've actually wanted to do and I have a following. And so like that, that's going to be how I kind of shoot my shot.
And yeah, I, I don't know what will happen, but like, I am getting that urge that I need to shoot bigger. I don't know.
Yeah, I mean, the other reason to do it is because if you think video editing is decent now, and it just kind of happened recently, I assume. I mean, I haven't used the models for video editing, but I've kind of seen people use it. I mean, just imagine what it's going to be like in a year.
I mean, it's going to be on another level. I mean, to put this into perspective, like think about where... I said that there was 10 months between Opus.
4 .1 and opus 5 .5 i don't know if you remember opus 4 .1 but like i remember my benchmark for opus 4 .1 was flappy bird like i was like okay in how many prompts could i create a semi non -deformed version of flappy bird and it was like six or seven prompts on ios and now with opus 5 .5 in one prompt i had a working call of duty game that i played on stream on stream multiplayer 3d 10 months and so like if you in and so like with that same slope might be happening right now with with video editing i think it is highly yeah i think it is yeah and and it's happening but it's it's just you know it's happening with robotics that's why all the roboticists are freaked out because they just they just watch someone get it use clod or astra or whatever and and control a robotic arm and like write a sentence or draw a drawing And they all flipped out because, and the reason, and it's always important when the experts in the industry freak out because they know, because they know that if it can do that, what's the call of duty?
I mean, that was Flappy Bird. That's like a Flappy Bird moment in robotics. Like, what is that going to be in a year?
I mean, next year, it's going to be great. Like next year, my prediction is that we're going to have, you're going to be able to write, you never, I mean, just to marry some of the things that we were talking about together, you're talking about everything is code. the robot the robots are going to be are being controlled by code you're going to like write a code that says move arm to pick up glass of water and then the arm will just go and do it and so i think that next year what we might see is even a proliferation of people getting into robotics that really had no business being in robotics like ordinarily they would have been like i can't do robotics well now you can and because being a software engineer is dead so that doesn't so writing code doesn't really matter so to be a roboticist i don't know it might be it might be more focused on like creativity and ideas and finding problems and doing interesting things with robots and like as opposed to all the building cool products yeah like exactly yeah like yeah like robots you still need to build like i saw i saw on twitter like a little robotic lamp
and that like could like look around uh it like sat on your desk and you could talk to it and you could use any model that you wanted i was like i want that i mean it was just like it reminded me of the pixar lamp um yeah and i think it's going to get really weird but i mean everything that we've been talking about through this conversation it's going to get really fun i'm nothing but optimistic yeah it's it's so much fun like you can but you can see the expansion you know just just to people watching this i hope you can like see the expansion of opportunity, you know, like 99 .99 % of the world couldn't do anything with robots or video editing.
Yeah. It's insane. And so you can create your own, you know, funny enough, I've been talking to some creators recently who have like big followings on social media and they're like, and we were talking, we're like, should we hire like a head of gaming?
And cause we were talking about like a lot of. Creating content is about world building, right? That's why you have like the similar fonts and you have your brand.
And a lot of people make PDFs. You know, Gary Vee has these purple and yellow PDFs that he sends out to people. That's what he uses for lead magnets.
And like he used to have to have a big team to construct these decks. And now he probably uses AI to construct these decks. I'm like, wait, we can create video games.
Like why would people not start making games as lead magnets or like hosting an entire video game world that... your followers can play. And like, and so I'm starting to like realize that like scaling your ambition might just be like the absolute best thing to do because anyone with a following can kind of create like a little media empire, right?
You can create films easier. You can create video games easier. There's so many other surface areas that you can, that you can win in.
And I think that's, that's just where this is all going. Yep. A hundred percent agree.
Yeah. Yeah. Anyway, this has been amazing.
I really appreciate you coming on the podcast. Is there any final thoughts? Any final thoughts or where we can find you or what we should try out with Particle?
Yeah, I mean, we'd love for people to try out our new model, Particle 1. We're already training the next one. I mean, we'd love people to use this.
I mean, the main thing we're trying to do is just also just sort of flatten things. We're not very excited about there being kind of like two big frontier labs that kind of tax humanity on intelligence. And so to the degree that we can just like really redistribute this to as many people as possible so that people can like scale their mission to all kinds of crazy things.
Because right now the malls are very expensive. You know, most people cannot afford two $200 max plans like you and me. And so I just think, yeah, I think that's a big part of what we hope to do.
So yeah, try us out. Very cool. Yeah, I use Vercel's AI Gateway.
Are you guys on the Vercel AI Gateway? We're on OpenRouter, Vercel, Cloudflare. That's probably like five more.
Yeah. Amazing. I will check it out right after this episode.
I appreciate it. All right. Thanks so much, Riley.
The Hook
The bait, then the rug-pull.
The cold open stacks the guest's most alarming lines, software engineering is dead and experts freaking out because they know, before the host reveals the real hook: a billion-dollar founder ran his own tests and found the newest Claude model only 10 to 20 percent better than one that cost thirty times less.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A 20-item weekly rundown of AI agent news: Claude Dashboards and Motion, GrokBot routing to Opus 5.5, ChatGPT's inline mini-apps, Nano Banana 2.1, and Adobe's open-source clones.
TypeSafe's new model can't write a sentence. It just makes fast, cheap decisions, and a growing list of builders are already wiring it into email triage, trading bots, and self-driving demos.