Lohithsai Yadalachanchu

Bob the builder, but my name is not bob

[Lohithsai Yadalachanchu @LohithsaiYC]

frog

[Yuchen Jin @Yuchenj_UW]

UC Berkeley open-sourced FreeToken. Wild results:

A single RTX PRO 6000 runs the 753B GLM-5.2 at 14.9 tok/s!

An 8GB RTX 4060 laptop (~$1,000) runs Qwen3.6-35B at 39.3 tok/s!

FreeToken is 2–4x faster than Ollama across consumer GPUs. Local AI inference is getting very real. Great work by @Andy_ShuoYang and UC Berkeley Sky Lab!

[Lohithsai Yadalachanchu @LohithsaiYC]

stripe was the og openrouter?

[Sean Mullaney @seanmullaney]

Why would @Stripe pay a reported $7bn for @OpenRouter? I led the team that built local payments at Stripe, alongside many of the @SeapointHQ team. Here's my inside take, and it's not just about tokens and AI.

My team at Stripe and I built and ran 40+ local payment methods worldwide, processing over $100bn a year. Each one took around a year of partnership negotiations, engineering, integration, testing and certification. It took a lot of clever design to abstract away all that complexity and variation behind a single API, and then operate it at high scale and >99.99% reliability.

In a single global checkout, Stripe would route each transaction to the local payment method most likely to convert that shopper into a customer. A merchant integrated once, sold globally, and made more money. The integrations were the moat. The routing was the product.

Stripe's own framing is about tokens. The new currency of AI, scarce compute, helping businesses spend it efficiently. That's the strategic story, and it's true.

I think the real reason is simpler. OpenRouter is structurally the same company as Stripe, just with models instead of payment rails. Both need clean abstractions, hundreds of integrations, routing and reliability at scale. And there is no shortcut to the hard work of building this. But whoever does the work to simplify that complexity will win the developers and get paid a premium for it, because developers will always pay to move fast and never think about what's underneath. That's what developers pay Stripe for.

The other half of this is the customer. Stripe won developers with the first router. OpenRouter's customers are developers making the same call for the same reasons. So this isn't Stripe buying into a new market, it's a second router sold to the buyer it already has.

And it genuinely works. We use OpenRouter on Seapoint's transaction enrichment and invoice reconciliation pipelines, running hundreds of A/B tests to find the model, prompt and workflow combinations that score best against our own eval datasets. It moved several workflows off Claude onto Gemini Flash, which was faster, cheaper and higher performing. Often the winner isn't one model at all, it's two combined, each doing the part it's good at.

I also use it daily as a panel of experts. Half a dozen reasoning models critiquing my thinking and arguing with each other beats one model agreeing with me.

Congrats to both teams. Very excited to see what you build together. Seapoint are very happy Stripe and OpenRouter customers!

If you're running models in production, are you routing across several or committed to one?

[Lohithsai Yadalachanchu @LohithsaiYC]

@zachmoskow @tryramp maybe looking to stripe gives some answers: https://t.co/rqDOo9GCMK

[Eric Newcomer @EricNewcomer]

Here's Stripe's letter to investors explaining its acquisition of OpenRouter (LEAKED)

[Lohithsai Yadalachanchu @LohithsaiYC]

@zachmoskow @tryramp why is a fintech company doing model routing???

[Lohithsai Yadalachanchu @LohithsaiYC]

@hanghuang_ @insforge 😭

[Lohithsai Yadalachanchu @LohithsaiYC]

my favorite part

[Lohithsai Yadalachanchu @LohithsaiYC]

A weird API request from a "user" leads down a rabbit hole involving belarus😭
https://t.co/IfkPbuJWHo

[jietang @jietang]

Thoughts About Scaling Law

Scaling, but not only of parameters. Every model release now ends with the same question: how many parameters? It isn't a question that can be answered on its own. Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions.

The field learned this the hard way. Kaplan et al. (2020) fit an exponent that told everyone to grow parameters faster than data — roughly 2.7:1 — and the industry complied: GPT-3, Gopher, MT-NLG. Hoffmann et al. (2022) redid the experiment across four hundred models and found the compute-optimal split is closer to 20 tokens per parameter, and that with sufficient compute the two should grow at the same rate rather than drifting apart. The error in the earlier fit compounded with every order of magnitude of compute, which is why the largest models of that generation were the most misallocated. The trillion-parameter round was, in retrospect, a detour the whole field took together and then reversed.

Chinchilla wasn't the end either. It optimized training compute for models that would be trained once and evaluated. Today a model is called billions of times a day and inference dominates lifetime cost. Put inference into the objective and the optimum moves toward smaller models trained far longer — deliberate over-training, which is what Llama-2-7B and Gemma-2-9B were doing at roughly 290 and 889 tokens per parameter.

Sparsity moved the target again. In a MoE model two quantities have to be kept apart: total parameters govern roughly how much the model can hold — knowledge, facts, the long tail — while activated parameters and effective depth govern roughly how far it can think, how many steps of a causal chain it can carry before it comes apart. A dense 20:1 ratio does not transfer. And the ratio isn't a single number at all: Roberts et al. (2025) find the optimal tokens-per-parameter is task-dependent, with memorization favoring more parameters and reasoning favoring more data. Follow-up work on MoE observes that at fixed TPP, pushing total parameters higher actually degrades reasoning, while activating more experts reliably helps it.

This matters for what we are building toward. Finding a vulnerability is not a retrieval problem. It doesn't come from having memorized more CVEs; it comes from carrying a twenty-step chain of inference to the end without losing the thread. That capability does not live in total parameter count.

Which brings us to this release. Total parameters appear to matter up to a threshold — enough to hold the world — after which additional capability comes from scaling elsewhere: effective depth per forward pass, and above all post-training. GLM-5.3 is our controlled experiment on that claim. Same base, same architecture, same total and activated parameters as GLM-5.2. One month of scaling long-horizon environments and RL. The gains are not marginal. Well, scaling has more than one dial. We turned the post-training one this time because it had the most slack left in it — not because the others are finished. Base model size, pretraining data, compute spent per forward pass: all of them are still on the table, and we will come back to each. What this experiment taught us is that the dials do not have to be turned together, and that the one worth turning next is rarely the one that was worth turning last. We are not done scaling. Next time, maybe mid-training, pre-training, and even more.

[Lohithsai Yadalachanchu @LohithsaiYC]

@mwseibel Github: https://t.co/IppaS6mJ4x

[Lohithsai Yadalachanchu @LohithsaiYC]

@mwseibel I built “policylab” at a hackathon. It takes a county, seeds a population of llms to match the county demographics using census data, and allows you to make your poll questions, which the simulated personas vote through.
1 min demo: https://t.co/eo3SgpWigM

[Lohithsai Yadalachanchu @LohithsaiYC]

@jia_seed what are u coaching?

[Lohithsai Yadalachanchu @LohithsaiYC]

“Oh, The Places You’ll Go!” is a good book to read if you’re going off the beaten path

[Lohithsai Yadalachanchu @LohithsaiYC]

@blwiertz Mfs be cooking up any excuse not to take public transport 😭

[Benjamin Bekken ⭐️ @BenjaminBekken]

Some of the most cracked founders in SF right now are building Orizon.

Lohith and Boris are cutting AI agent inference costs dramatically with a gateway that learns from production behavior.

If you're burning cash on inference, go talk to them!🔥

[Lohithsai Yadalachanchu @LohithsaiYC]

State of the world rn

[Lohithsai Yadalachanchu @LohithsaiYC]

Super cool blog post about elevator algorithms
https://t.co/U4G5ZRZZlw

[Lohithsai Yadalachanchu @LohithsaiYC]

This is the pacific ocean. It's also the furthest I've been from "home" and it's got me thinking.
We live in an era where a 20 smth year old can go to the other side of the world and pitch their idea of how the world should be. And if they convince the right people, they actually can. For most of human history this was not the case, and even today there's not many places this holds.
I think what makes these places special is the density of people who are brave enough think (and share that they do) from first principles, instead of simply following the herd without thinking why.

[Lohithsai Yadalachanchu @LohithsaiYC]

Loved chatting about ai safety and VendingBench with @lukaspet from @andonlabs yesterday at YC startup school!

[Lohithsai Yadalachanchu @LohithsaiYC]

What is this merch 😭