VP of Eng @ Chronicle Studios | ex-Google
SF Bay Area
Joined February 2024
- Tweets1.8K
- Following477
- Followers321
- Likes2.5K
Pinned Tweet
For the entire history of engineering, we designed organizations around code being expensive. Code is becoming cheap, and the org chart hasn’t caught up yet.
Most of the conversation is still about productivity: faster coding, more output, fewer bottlenecks.
The far more interesting question is how much more of the product one engineer can responsibly own, and what this means for teams, architecture, and management.
I explored what that might look like here: medium.com/@jengas/ai-is-cha…
It's never been easier to do something impressive with AI. Getting it to actually work inside a real company? That's another story.
I was chatting with a few other executives during a16z Tech Week in SF, and almost every conversation eventually came around to AI. Not just the technology itself, but how we transform our organizations, our products, and the way we operate.
And one thing has become really, really obvious: this is ridiculously hard.
Working with AI can feel like magic. Type in an idea and out comes an essay, an image, a video, or hell, even a video game you can play.
But that experience tends to overshadow just how steep the learning curve actually is. Especially when you're trying to build real products, for real customers, inside real businesses.
I keep seeing companies wrestling with the same problems I was struggling with a couple of years ago.
It's incredibly tempting to start with the technology. Build something impressive, show everyone what it can do, and then figure out how to get people to use it.
I've learned the hard way that taking the opposite approach works a whole lot better.
Start with the people doing the work. Find a real bottleneck, understand the economics, and put someone who deeply understands AI alongside the people who actually understand the problem.
Those people know what good looks like. They understand the nuances of the work, which mistakes actually matter, and what separates a genuinely useful result from something that just looks impressive.
That's how you build meaningful evaluations. Turn their expertise into a way to measure quality, then get something into production and test it against reality.
For example, an AI model predicting what an audience will respond to is interesting. Actually publishing content, measuring the response, and finding out whether the model was right? That's where things get useful.
Cost, reliability, quality, actual business outcomes. Measure the hell out of it, learn from the results, and keep improving.
Otherwise, how do you even know whether your AI is doing a good job?
Then, and this is the really important part, make yourself unnecessary.
Give the team ownership. Help them develop the skills, tools, and judgment to keep improving what you've built together. Take what's reusable, apply it to the next problem, and repeat.
The goal here isn't to build AI capabilities. It's to build an organization that knows how to use them.
This little experiment is starting to get a bit out of hand.
Last week I showed three AI models running entirely inside a web browser, letting me command an RTS game with my voice.
No cloud inference, no API calls.
Over the weekend, I turned my local GPU cluster loose on the game's design document. A gauntlet of coding agents worked through the remaining systems and features, building out what was still missing.
Now it's starting to look like a real game.
Sure, I'm using AI to build the game, and AI is running inside the game. But this isn't just about development speed. The economics of building and operating games have changed, wildly.
Of course, none of that matters if the game isn't fun.
Challenge accepted.
2013 was the last time I hand-optimized assembly
now I know I'll never have to do it again
life is good
I have three generative AI models running inside a web browser. On my phone. Not a web page calling AI APIs.
The actual models are running on the page inside the browser.
I replaced Gemini Live in my RTS game prototype with an entirely on-device AI Commander.
One model listens to my voice. A second, fine-tuned 350M parameter model understands my orders. A third speaks back. The game's own systems handle the strategy and execution.
I can tell it to build defenses, move units, ask about the battlefield, or change my mind mid-conversation. All without a single inference request going to a server.
It works on my Mac. It works on my Pixel 11. The decisions take roughly half a second, with no measurable impact on the game's frame rate.
But here's where this gets really interesting: in a game built around generative AI, where using the AI is native to the experience, I just eliminated the cost of running that AI for every player.
Normally, the more people play, the more they interact with the AI, the more the inference costs.
Not anymore.
The models ship with the game. Every player brings their own compute. They can interact with the AI as much as they want, and the inference cost stays at zero.
No AI subscriptions. No API keys. No usage limits. No inference bill for me or the player.
And because it's all running in a browser, I can deliver that experience across desktops and phones, anywhere really that has enough compute to run the game, all without building separate native applications.
This fundamentally changes the unit economics of publishing an AI-native game. But I think the implications go a lot further than games.
What happens when generative AI stops being a service you rent, and becomes software you can simply ship?
For the technical folks curious how this works:
Three open-weight models, all running inside the browser:
- Moonshine Small Streaming (166 MB): Speech recognition with voice activity detection, running on WebAssembly.
- LFM2.5-350M (229 MB): Fine-tuned to translate natural language into game commands. Runs through wllama/llama.cpp on the CPU, with grammar-constrained decoding.
- Kokoro-82M (~109 MB): Generates the Commander's spoken responses.
The language model was trained on ~45K examples, augmented with synthetic speech and actual transcription errors. It scored 100% on 160 held-out text commands and 98.8% on speech-to-command tests.
The model doesn't make gameplay decisions. It interprets intent and calls into the game's existing deterministic planning and execution systems, which enforce validation, budgets, and permissions.
My latest video experiment reminded me that there's a lot more to making a good AI video than generating good footage.
I spent some time working with Google Veo this weekend. I was able to dial in the look, but the fight choreography? Terrible.
Rather than burning through a bunch of generations hoping to get lucky, I tried to get clever with editing instead.
Turned the same 3 shots into 8 cuts. Better pacing, better timing, remastered sound.
And it worked.
Is it perfect? Not even close. But look at the before & after.
Sometimes making something better isn't about throwing more compute at it. It's about getting more out of what you've already got.
All footage generated with Veo. Editing and production handled by my custom agent and toolchain, powered by Opus 5.5 in Claude Code.
I put Google's new Gemini 4 Argon through my one-shot Asteroids test, and you can see the result for yourself.
The initial game was good, but what I really liked was the visual QA. Argon identified and fixed z-fighting, broken bloom effects, blurry planetary poles, and pixelated storms. Not by guessing, but by playing the game in the browser, tracking down the actual rendering problems, and fixing them.
Ended up with 81 shaders at 60 FPS in the browser.
Still a few rough edges, but the level of engineering and visual polish coming out of these models is getting ridiculous.
Your next favorite movie, game, or creator probably already exists. You just haven't discovered them yet. And the person who made it probably has no idea you'd love it.
That's a pretty crazy problem when you think about it.
We have more ways to create and distribute content than ever before, yet finding the right audience is still largely a guessing game.
At @Chronicle_tech, we've been working on changing that.
We started by testing tens of thousands of real-world video and audience pairings. One surprising discovery? How similar two pieces of content are tells you almost nothinge about whether their audiences will respond the same way.
So we built UBQT, our own audience intelligence model, and put it head-to-head against GPT, Claude, and Gemini.
The results were pretty wild.
Here's my favorite example: all three LLMs recommended mainstream cooking channels for a baking video. Seems reasonable, right? Except only 1.6–3.4% of those audiences chose to watch.
Our model recommended a mystery/conspiracy channel instead, and 40% chose to watch.
Across the broader benchmark, audiences discovered by UBQT watched 2.5x as often as those recommended by the best-performing general-purpose LLM.
Turns out understanding what something is about and understanding who will actually love it are two very different problems.
I've spent most of my career building games and entertainment. Imagine being able to discover an audience for your weird, risky creative idea before spending a fortune bringing it to life.
That's not just better marketing. It fundamentally changes which creative bets become possible.
Incredibly proud of what we're building at Chronicle, and the team behind this research.
Josh English retweeted
Every brand wants fans. Now there’s an easier way to find them...
Meet UBQT, our first frontier model for audience forecasting. Brands and studios can now predict how niche audiences will respond to content before they spend.
Check out our new and improved website, linked linked down below!
The fans are out there. Let’s go find them 🚀
The entertainment industry is pretty good at figuring out what audiences loved yesterday. Predicting what they'll love tomorrow? Not so much.
That's one of the big problems we're tackling at @Chronicle_tech
We're building AI that understands audience taste, predicts demand, and discovers communities of fans that nobody knew were there.
The idea is simple: great entertainment shouldn't have to rely on luck to find its audience.
We just launched our brand-new website, and I'm incredibly proud of what this team is building.
Take a look: chronicle.ai/
Another weekend, another experiment. I took my free Google AI Pro subscription, an AI agent, and an idea for a music video. About an hour later, it handed back this.
Lyria 3.5 composed the song, Nano Banana created the look, Veo 3.1 brought it to life, and Omni fixed shots that didn't work.
The agent handled everything else: continuity, lip sync, editing, color grading, typography. When it needed a tool, it wrote one.
No shot list. No editing. And I'm genuinely blown away by how good it is.
With all the attention on Suno and Seedance, I was honestly wasn't expecting this level of quality from Google's models.
And the whole thing cost me basically nothing.
It makes me wonder how far we are from anyone being able to tell any story they can imagine.
I had a blast at the a16z Speedrun AI Faire event yesterday!
From humanoid robots to 3D worlds, I got to see a whole lot of incredibly talented people showing off what they've been building.
What struck me wasn't any one demo, but the sheer variety of things people are trying.
For most of the history of technology, building something ambitious was expensive enough that an awful lot of good ideas never got a chance.
AI is changing that equation at an extraordinary pace.
And the exciting part isn't just that we can build the same things faster and cheaper.
It's all the weird, ambitious, wonderful things people can finally afford to try.
What a ridiculously fun time to be a builder!
AI has broken hiring.
We’re evaluating people using information that made perfect sense as technology and the world around us evolved at a predictable rate.
Historically, we got away with using the past as a proxy for the present. If someone had spent twenty years becoming great at something, that told you quite a lot. The tools changed, but the job hadn’t, so accrued experience still compounded cleanly.
AI broke that assumption.
Someone can have twenty years of software experience and still be almost completely disconnected from how software is being built today.
A company can spend years living off a reputation earned by products it no longer knows how to make.
And there are far too many people still treating AI like it’s the ChatGPT they tried once in 2023.
This is one thing the startup world gets right. If someone nobody has heard of shows up tomorrow and builds something extraordinary, people pay attention.
Proof is greater than pedigree.
Traditional hiring still leans heavily on titles, logos, credentials and years of experience when it has never been easier to see what someone can actually do.
Experience and reputation still matter, but only insofar as they remain connected to reality.
The faster things change, the more weight I put on what you’re doing now.
And if your opinion of AI was formed three years ago, you don’t have a current opinion about AI.
You have a memory.
Google Labs just launched Playground, so naturally I gave it a ridiculous test.
I asked it to build a 3D chess game where the pieces battle each other like Wizard’s Chess from Harry Potter, but set in a Halo-like sci-fi universe.
One prompt.
It understood what I meant: not a clone of Harry Potter, not a copy of Halo, but take the underlying ideas from both and make something new.
So it built a sci-fi 3D chess game where the pieces actually fight each other, ran automated QA against what it had made, fixed the problems it found, and gave me something I could publish and play.
I never entered the development loop.
And I did the whole thing from my phone, riding in a car on the way to an event.
As someone who has spent most of his career making games, that’s the part that feels wild.
Historically, having a game idea and actually making a game were two very different things.
You needed programmers, artists, designers, tools, an engine, a production pipeline, and enough technical knowledge to make all of those pieces work together.
Playground takes the game you have in your head and makes it.
Then you can play it.
Obviously this doesn’t suddenly make everyone a great game designer. Making something playable and making something good are very different problems.
But it dramatically expands the number of people who can try.
It moves the bottleneck from knowing how to build a game to knowing what good looks like.
That matters.
Millions of people have game ideas and absolutely no idea how to make a video game.
Now they can.
I gave GPT-6.1 Sol a one-shot test: build an Asteroids game on the Bitmagic GDK.
It worked.
That was the least interesting part.
The first version compiled, passed its tests, and was technically complete.
It also sucked.
The controls were wrong. The weapons felt bad. The art was weak. The game had lost most of what made the original good.
So the real test became something else: could the model recognize the difference between working and good, then close that gap?
Turns out, yes it can.
Bitmagic made the engine, tooling, assets and publishing accessible enough for the model to operate entirely inside the platform.
The AI model did an impressive amount of real gamedev, but the bottleneck wasn’t code.
It was taste.
The model could make the game work.
Making it fun required knowing what good looked like.
