dean

@MyriadVC

[noah @noahlofq]

added terminal support as well to Light Mac app now, so it can do stuff across CAD, Blender. At this point I'm using it to automate a lot of my own workflows, using only my voice.

[noah @noahlofq]

for forty years, you operated the machine.
in this demo, it's reversed. i talk, my mac works. acts across my apps, remembers my whole day.
fully on your machine, or in the cloud. your choice.

[Tomas Hernando Kofman @tomas_hk]

I'm giving a talk today about intelligent model routing at @aiDotEngineer World's Fair at 2:50pm.

I'll cover various model routing techniques and the specific approach we use at Not Diamond to achieve 30%+ savings on coding agents with no loss in quality. Come check it out!

[dean @DeanMai]

"the spice must flow." this is why the debate about token costs, token efficiency, and model routing matters. in a world of rising token prices = possible rationing, the product that can swap in cheaper or open-weight models keeps serving customers. https://t.co/w23MQRmAkB

[Matthew Berman @MatthewBerman]

If you want to know how to achieve this, talk to @tomas_hk

This is what his incredible team is building at NotDiamond.

[Tomas Hernando Kofman @tomas_hk]

Enterprises are increasingly looking to route between closed and open source models to manage exploding inference costs.

Over the past eighteen months, I have seen the attitudes of F500 leaders shift when it comes to open source models, particularly Chinese models. The growing distribution of tasks that can be handled by these models together with the urgent need to reduce AI spending has organizations looking to intelligently route between powerful frontier models and leading oss models to keep the productivity gains from coding agents without sending every task to the most expensive model by default.

That's a very hard thing to do.

- Naive approaches to routing backfire
- Models change constantly
- Developers choose quality when given the choice
- Education does not solve model selection
- Token usage does not translate linearly to ROI

This is why we're building infrastructure that determines which model should handle each request, so enterprises can preserve frontier-level quality without paying frontier prices every time.

Across enterprise pilots, we help orgs with thousands of engineers achieve 30%+ lower inference costs while maintaining frontier-model quality.

Good article from @theinformation's @LauraBratton5 on what we’re seeing across the world’s largest engineering teams:

[Tomas Hernando Kofman @tomas_hk]

Everyone is looking into model routing as companies blow through their inference budgets. But most engineering leaders still don't actually know what model routing is, or how it's different from an AI gateway.

The two are easy to confuse because both are critical multi-model infrastructure. But they represent extremely different problems to solve.

A gateway provides access to various models. An intelligent router automatically recommends the best model at the lowest cost for each request. The router sits between your agent and your gateway, returning a model recommendation for each request which is then executed through your gateway.

When we started working on intelligent model routing back in 2023, we made a decision to not be a gateway so that we could a) integrate with any gateway, and b) stay laser-focused on solving routing. While gateways are fundamentally an infrastructure and engineering challenge, model routing also spans research, ML, and data science, and it requires constant reinvention as new models are released every week.

I think everyone should get sharper on the distinction between routers and gateways, so we've broken down the differences here: https://t.co/HyvPThQp2x

[Tzafon @tzafon_company]

Mac environments are now live in Lightcone. Run the fastest computer use model on Earth on macOS - Xcode, Safari, native apps. Sign up for access below.

[dean @DeanMai]

teams that build routing in-house underestimate it; they treat it as a classifier you train once, but you have to re-solve it as models shift and can’t tell you’re losing money without a per-task quality signal most teams don’t have. exactly the problem for specialized infra.

[Tomas Hernando Kofman @tomas_hk]

My thoughts on the future of model routing and AI:
- We have not even scratched the surface of runaway inference costs
- Solving this requires intelligent model routing, especially as the inference landscape continues fragmenting. This is a *hard* problem.
- Naive solutions (turn-based routing, session routing) fail; routing successfully involves managing multiple cost surface areas in concert.
- Getting routing right means a more diverse market of providers, more power for consumers, reduced ecological impact, and improved effectiveness.

More in the full essay:

[Tomas Hernando Kofman @tomas_hk]

My thoughts on the future of model routing and AI:
- We have not even scratched the surface of runaway inference costs
- Solving this requires intelligent model routing, especially as the inference landscape continues fragmenting. This is a *hard* problem.
- Naive solutions (turn-based routing, session routing) fail; routing successfully involves managing multiple cost surface areas in concert.
- Getting routing right means a more diverse market of providers, more power for consumers, reduced ecological impact, and improved effectiveness.

More in the full essay:

[dean @DeanMai]

stealth reallocation. Al lowers the viability threshold for tiny hyper-specific ventures + quietly absorbing the long tail of uneconomic micro-tasks that big companies have always ignored or outsourced at high friction (without showing up as job loss). https://t.co/jDeGlXor0l

[Tzafon @tzafon_company]

Builder Weekend starts today at 5PM PT and runs through midnight PT on Sunday.

Northstar CUA Fast, the world's fastest model for computer use agents — 90% off all weekend.

[noah @noahlofq]

so Lightcone is the fastest computer use agent in the world and watching it do the boring stuff this fast is genuinely kind of insane. and it's only getting faster and more reliable.

[Tomas Hernando Kofman @tomas_hk]

Grok 4.3 is live in the Not Diamond router.

@xai's strongest model built for long horizon coding tasks, tool calling, and instruction following, with a very clean quality/cost-efficiency profile. You can now route to it automatically alongside frontier models from every major lab through Not Diamond's model router.

Thanks to Frank Ryan and the xAI team for the partnership.

[Tomas Hernando Kofman @tomas_hk]

On the @theallinpod, @Benioff describes why routing is the next layer of enterprise AI infra—and how it will save billions of dollars.

We've been building exactly this at @notdiamond for two years. Largest vendor of intelligent routing in the world. @Benioff, we should chat!

[Tzafon @tzafon_company]

We just ran the world's largest Computer Use hackathon with @nvidia, @usekernel, and @sfcompute.
20 teams, pushing what agents can do on computers.

Here's what they built 👇

[Kabir Nagrecha @KabirNagrecha]

Today we’re announcing Tessera’s $60M Series A led by @a16z.
Enterprise transformation is broken – years-long timelines, massive cost, and high failure rates. Tessera is rebuilding it with AI, delivering in weeks what used to take years.

We’re hiring engineers, researchers, and operators who want to help rebuild how the world’s most important systems evolve. If that’s you, reach out.

[Tomas Hernando Kofman @tomas_hk]

Copilot is increasing the cost of Anthropic models by 9x. The era of subsidized inference is coming to an end.

Coding agent spend 10x'd in the last six months because of usage. Now usage is going to 10x again on top of 10x higher inference costs. If you think your enterprise coding agent bills are big now, it's time to brace for impact.

What can you do to prepare?
- Analyze your usage and spend patterns to understand the source of the problem
- Renegotiate your enterprise contracts
- Talk to me about intelligent model routing

[Tzafon @tzafon_company]

Today we're releasing the Lightcone SDK – the fastest path from idea to working computer automation, in Python or TypeScript.

The most powerful SDK for computer use.

[oleander @OleanderHQ]

We launched our skills repo today with context about how to write lineage compliant PySpark scripts and work with Iceberg seamlessly. Repo is available below. https://t.co/1MrqtIMcQn