Jorge Galarza

Engineering , AI .

[tanya @Tanya_Sabrinaaa]

wait until he hears what motivates investors

[Mario Zechner @badlogicgames]

recommended reading. similar observations on my end, albeit at much smaller scale.

[herdr @herdrdev]

this is great! but don't leave claudes alone :(

in herdr any agent can talk to any agent.

no protocol, no MCP, no setup. it's just terminals.

[Jorge Galarza @jorsisu]

@finkd Where is GPT-5.6 Sol ?

[Matt Pocock @mattpocockuk]

mattpocock/skills v1.2 is out!

We're now the 19th most-starred repo of all time. 13.5m downloads on skills​.sh. Thanks for your support!

Here's what's new:

- Docs: the community's biggest ask. Every skill documented, with explanations of the main flows + troubleshooting
- Claude Plugin: install via Claude's official marketplace
- Codex Support: full Codex support via agents/openai.yaml files

Updated Skills:

- /grilling now asks you questions in rounds, not one-by-one
- /prototype now uses HTML instead of a TUI for building logic prototypes - easier to share and far richer
- /writing-for-agents renamed from /writing-great-skills, use it for ANYTHING your agents read (AGENTS.md, system prompts, docs)

New Skills:

- /wizard: tired of provisioning infra? Get your agent to build you a TUI to walk you through it
- /to-questionnaire: hit a grilling question you can't answer? Turn the session into a doc you can walk through on a call with a colleague
- /wait-what: no idea what the model said? Refocus it in your domain language and simplify with ASD-STE100

Full changelog + docs below. Video soon!

[Jorge Galarza @jorsisu]

“Models are the engine, and Fireworks is the race team (prep, pit crew, strategy, and fine tuning of the car). You can have the best engine, but without the best race team, you are not winning. 🥇

[Paul @P_Bonnet]

You probably used @FireworksAI_HQ this week. The cool part is that you just did not know you did. 🎆

Open-source models are free, sure... but the hard part is tailoring them so they perform best on your own use cases. And making them run fast, cheaply, and reliably in production. This is the mega hard work.

*Fireworks is the layer that does this exact work:* it sits above the chips, but below the model, abstracting away the complexity of deploying and optimising open models.

As a result, they already crossed >$1bn in run-rate revenue, up 5x YoY... and they are accelerating at scale!

20VC invested $10m in the most recent round. Here is why @HarryStebbings and I at @20vcFund were excited to do so:

*First, the team:* the company was founded in 2022 by @lqiao and six co-founders, largely the team that built and ran PyTorch (the framework nearly every AI model on earth is trained on) at Meta. That is truly a unique right to win in that space: if anyone knows how to make models run fast on GPUs, it is the Fireworks team.

*Second, what they do:* the best analogy I can come up with is Formula 1. Models are the engine, and Fireworks is the race team (prep, pit crew, strategy, and fine tuning of the car). You can have the best engine, but without the best race team, you are not winning. 🥇

*So how does it work in practice?* Two options:

(1) You just pick an open model off the shelf (DeepSeek, Kimi, Llama, gpt-oss...) from the 400+ already running on their platform, you call the API, and you pay per token. You do not need to rent GPUs or manage clusters: their hand-written GPU code (down to the kernel level!) runs those models. And they do so faster and cheaper than you could yourself.

Or, and most importantly (2) when off-the-shelf is not enough, they help you fine-tune an open model on your own data. With that: a specialist model, at a fraction of the cost, will beat a frontier generalist on your specific tasks. That approach is available self serve on their platform, or, for large customers - through FDEs who work directly with them to optimise models for their specific workloads.

That is, in part, how @cursor_ai built its Composer model: on Fireworks. Similarly, @harvey achieved above frontier performance at 10% of the cost (yes, that's mindblowing) working hand in hand with the Fireworks AI teams.

*And why will they win, structurally:*

(a) inference is where the money is going: ~90% of a model's lifetime compute is spent running it (as opposed to training it). And agentic AI is a driving force behind the explosion in token volumes ==> the tide rises fastest where Fireworks sits, and I think this mega-trend will continue working in their favour

(b) open models have reached near-frontier quality ==> every CFO is starting to worry about their closed-model API bill. And now Fireworks enables them to off-ramp from that exponential cost scaling: they can continue performing at frontier level (or above!), but at 10% of the cost

(c) speed: per @ArtificialAnlys (independent benchmarks), Fireworks is consistently the fastest GPU-based provider on top open models. Only the custom-silicon players beat them on raw speed - but often at higher costs and with a fraction of the model coverage

(d) no lock-in: CEOs and CTOs do not want vendor lock-in given how fast the space moves. DeepSeek beats Kimi, Kimi beats Llama, Llama beats DeepSeek again... It does not matter: Fireworks earns on all of them. Model churn benefits Fireworks.

*I also reviewed 20 expert calls. Four observations stood out:*

1) one head of AI rated their support "14 out of 10". And no, that is not a typo! He added they would be "much worse off without them"
2) one customer moved a flagship AI feature off GPT-4o onto an open model on Fireworks: 70-80% cheaper... and usage went UP 20% (👋 Jevons Paradox)
3) a global creative-software giant scored them 9/10 head-to-head vs 7 for their closest rival, and is shifting more towards Fireworks as a result
4) a top-5 US healthcare company, asked to pick one inference provider to bet on: "I would go with Fireworks"

When Cursor writes your code, when Notion's AI summarises your meeting, when Upwork drafts a freelancer's proposal... they share the same engine underneath: Fireworks AI.

The winning layer may not be the model itself, but the infrastructure that makes every open model usable. Great infrastructure companies are invisible to users, but indispensable to developers!

Let's go Lin and Fireworks team 🚀

cc. Harry Stebbings, @Kieranleehill, @codorniou, @alexandre_dewez, @NiallKiely20VC

[herdr @herdrdev]

herdr 0.8.0 is out, and it's the first apache-2.0 release 🎉

not a feature-heavy one, it's groundwork for what's coming next. 55 fixes, 18 contributors, and the biggest performance work herdr has had.

what got removed, and why 👇

[OpenAI @OpenAI]

GPT-Live can listen while it speaks.

To make that feel natural at ChatGPT scale, we rebuilt the voice stack from client to model.

This new architecture keeps audio flowing continuously, so deeper reasoning and tool use don't interrupt the conversation.

[Matt Pocock @mattpocockuk]

I know I've reposted this before, but it's worth re-emphasizing.

If you want to get good at using AI, GET GOOD AT THE THING YOU'RE USING IT FOR

This is a great illustration of why:

[Guillermo Rauch @rauchg]

Open source agentic CRM built on https://t.co/99eEa13mZ3 and @nextjs.

Model-agnostic, self-hostable or serverlessly-deployed, multi-channel, and headless.

This is the way.

[James Russo @Rames_Jusso]

This video is all code built by chatting with @OpenAI Codex

This is why we built @HyperFrames_ anything on the web can be turned into a video, even webGPU code

Thanks to snowflow_demo https://t.co/KBQ7pCLwfx

[Jorge Galarza @jorsisu]

@dhh @nayibbukele if you want to use Bukele as a reference, you might want to listen to this first https://t.co/fnEVIjsRtV @dhh

[Jorge Galarza @jorsisu]

@joshpuckett @valiocon I'm in!
Great lead magnet, BTW

[Matt Pocock @mattpocockuk]

LLM's are notoriously bad at following negative instructions

That's why code review works so well, it turns all the negative instructions into positive ones:

[Addy Osmani @addyosmani]

Software quality now depends on the constraints you set around your agents.

When humans manually wrote most of the code we could look at the code itself for signs of quality. Is it clean? Is it thoughtful? Is it fast? Can another engineer understand it? Does it have tests?

Agents can now generate more code than people can read. When code generation scales beyond review, quality - checks for one or more of correctness, maintainability, security, performance etc - increasingly has to live somewhere else.

It moves into the harness, environment and operating system around the agent.

This can be the tests and deterministic checks that decide what the system is allowed to do (amongst others). Your constraints are what may eventually enable loops of agents to deliver production software reliably. They can include unit tests, property tests, acceptance tests, mutation testing and quality metrics.

This back-pressure lets the system resist bad work before it becomes somebody elses problem.

Set your constraints. They decide whether the code your agents generate is good enough to ship.

[Jorge Galarza @jorsisu]

@superlogical Isn't this https://t.co/79vSkkXdHR ?

[Jorge Galarza @jorsisu]

@mitchellh @superlogical Isn't this https://t.co/79vSkkXdHR ?

[Jorge Galarza @jorsisu]

@almonk @mitchellh @dizzyup Isn't this https://t.co/79vSkkXdHR ?

[shadcn @shadcn]

Launching Copper, a Mac app for capturing things you want to keep and prompts you want to try next while working with AI.

The more I use AI, the more I find myself collecting little things I don't want to lose.

You're in ChatGPT and think, "I'll need this later," but you don't want to stop what you're doing. Then you're in Claude. Cursor. Chrome. Back to ChatGPT.

Before long, you've got little things scattered everywhere. An answer you want to keep. A link. An idea. Or three follow-up prompts before the current one has even finished generating.

I had this problem, so I built Copper.

Copper combines the useful parts of a to-do list, a clipboard, and a scratchpad. It's built specifically for AI-assisted work. It sits next to where you work and is always one shortcut away. It works with all your AI apps, terminals and browsers.

Whenever I find something worth keeping, I capture it. If I think of a few prompts while the AI is still responding, I type them into Copper. Then I send them back into ChatGPT, Claude, or Cursor and check them off as I go.

Copper is local and private. It doesn't sync anything, doesn't collect anything, and doesn't need an account.

I've used Copper every day for the past few months. It has completely changed how I work.

If this feels like something you'd use, it's $39. One-time purchase.

[Jorge Galarza @jorsisu]

@mweinbach @FireworksAI_HQ Same