Charles

Conifer (YC S26) | @ Princeton

[Charles @charles_v11]

Pretty frustrating using Claude this month. Models seem slower, constantly hitting guardrails, and burning tokens. Planning is still superior, but definitely stresses the need for multi model setups

[Charles @charles_v11]

@MichaelJeffords LFMs are also really fast/cheep to fine tune. Always tradeoffs 👏

[Charles @charles_v11]

Kimi beats Fable and Sol in Code Arena!!

[Charles @charles_v11]

Kimi K3 Benchmarks - seriously where is for profit in 18 months

[Charles @charles_v11]

We just launched Juniper! Try it at https://t.co/QzHocqlXWU

[Conifer (YC S26) @coniferbuild]

Stop paying cloud prices for every AI request. We route 80% of your queries to your own hardware at no cost, using the cloud only for your most complex tasks. Available now @ https://t.co/Xj83Zhxf6r.

[Charles @charles_v11]

https://t.co/TGKUmS2VVH
Sharpest point imo: every time we call a capability "emergent," we're quietly admitting scaling laws can't predict what's coming. That's a strange foundation for trillion-dollar capex. From where I sit in inference infra, the efficiency curve is important; 22x isnt one off its a trend that's likely to dominate what the future of inference looks like.

[Charles @charles_v11]

Fresh essay on scaling laws.
Sara Hooker spent years inside DeepMind and Cohere watching scaling laws get treated as gospel. Then she wrote down the heresy:
Falcon 180B was state of the art in 2023. One year later it lost to a model 22x smaller.
The $700B question isn't whether scaling works. It's whether it's still the cheapest way to buy intelligence.
She was convinced enough to quit and build a company on the answer.

[Michael Jeffords @MichaelJeffords]

Excited to share that @coniferbuild is part of the @ycombinator S26 batch. Looking forward to getting to work with @gustaf!

Token costs are eye-watering. @charles_v11 and I burned through $13,000 in Claude credits in just five days, and that's a fraction of what every company transitioning to Al is facing.

But it's not just the cost. Al usage today is fragmented: five different models, three subscriptions, three API dashboards, and a drawer full of API keys. Every team is flipping between tabs, juggling providers, and paying full price for all of them.

Conifer replaces all of it with one interface. We're the inference gateway that routes every Al query to the cheapest model that can actually do the job. Built on top of our Typhoon engine, which allows local models to run faster and punch far above their weight, Conifer only reaches for the cloud on your hardest tasks. The result is a token bill >80% less.

We launch tomorrow, July 7th, on https://t.co/weRgjPtW0x

If you're part of a company spending thousands on Al every day, been holding off on the switch because of cost, or someone just trying to decrease their spend, we'd love to hear from you. Email [email protected] or reach out here on X.

[Loc @locbuilds]

I have to say, I agree. I've been reading a lot of research on LLM routing, and it has consistently shown that the majority of queries don't actually need a frontier model. Most of what teams send to Fable 5 would come back just as good from something far cheaper.

And even for the most difficult prompts, Anthropic literally just shipped Sonnet 5 last week. It's nearly on par with Opus 4.8 (beats it on some knowledge work benchmarks) at less than half the price. Defaulting everything to the "best" model in 2026 shouldn't be a requirement, and oftentimes (as you pointed it out), it can be costly.

[Charles @charles_v11]

So many great points, but the error-correction piece is the one people will underrate. LeCun's (1−ε)ⁿ doom assumes per-step errors are iid and absorbing, which they're not. Trained agents learn a restoring force back so long horizon behavior looks like a mean reverting walk. 👏

[bayes @bayeslord]

https://t.co/D6DjGcUFZI

[Conifer (YC S26) @coniferbuild]

You're wasting API tokens. A 14B model that runs on a laptop now solves 85% of AIME 2025

[Conifer (YC S26) @coniferbuild]

Conifer is an open-source runtime that makes local AI actually fast, private, and reliable.

Here's the problem we kept running into: running AI models on your own machine should be the obvious choice for anything private, your code, your documents, your data. But anyone who's tried knows the truth: the hard part isn't the model, it's everything around it. Setup. Storage. Quantization. Memory management.

Getting it to actually use your GPU properly. Splitting a model across multiple cards without it falling over. Keeping it stable when it runs for hours instead of minutes. Most people hit that wall and go back to the cloud. Which means their data leaves their machine, their costs scale forever, and their tool breaks the day someone else's servers have a bad day.

Conifer is the layer that handles all of it. It manages the model, setup, storage, quantization, and memory, and schedules the work across whatever hardware you actually have, so local inference just runs. The goal is simple: local AI that's genuinely competitive with the cloud, not a slower fallback you settle for. And we're getting there. In early benchmarks Conifer already beats llama.cpp on decode for some models (1.24x on Qwen, 1.13x on TinyLlama) and MLX on prefill (1.26x on Qwen). It's not a clean sweep yet, there are places we're still behind and tuning hard, but it's genuinely competitive, and the numbers are climbing before launch.

Conifer launches June 1st. Completely free and open source. Install it, point it at a model, and see what it does for yourself, no signup, no trial, no catch. We're opening beta access to the first 100 people on the waitlist before then. If you run models locally, or you've wanted to but the setup beats you, we'd love for you to try it and tell us what breaks.

Link for signups will be in the replies.

[Charles @charles_v11]

https://t.co/HDeh4ixykZ

[Charles @charles_v11]

CUDA: Why Modern GPUs and LLMs Depend on it

This article examines CUDA as the execution contract between software and GPU hardware

[Charles @charles_v11]

https://t.co/Ws7aJoBjdU

[Charles @charles_v11]

Peak FLOPs are a conditional truth

Tensor cores advertise absurd compute density, but those FLOPs only exist if data arrives perfectly.

This article breaks down what happens when that assumption breaks: memory coalescing, tiling, and warp scheduling become the real constraint.

[Charles @charles_v11]

Why ML Performance Is a Hardware–Software Problem

An examination of GPU execution for ML, focusing on how model structure, memory access patterns, and scheduling interact with hardware constraints during training and inference.

https://t.co/r1WjkdwpxR