I like to make cool stuff and watch Football | Community and Marketing @Chutes_ai | Decentralised AI Maxi
Joined January 2022
- Tweets1.2K
- Following193
- Followers278
- Likes2.1K
Pinned Tweet
AI is cheap the way Uber rides used to be cheap.
Investor money is eating the real cost to get you hooked.
The day the math has to work, your bill jumps and you get zero say in it...
followed the money so you don't have to👇
Fezicles.eth retweeted
When you send a prompt to Chutes with E2EE, it runs inside a hardware-attested TEE with end-to-end encryption. We can't read it, and neither can anyone who asks us for it.
Even holding the TLS certificate for chutes.ai, we don't see prompts or responses. If we were ever ordered to hand them over, all anyone would get is ciphertext.
The whole stack is open source, so you don't have to take our word for it.
Fezicles.eth retweeted
Introducing DecisionModels @decisionmodels_
I've spent the last four weeks benchmarking decision models (I run JevBench).
Today I'm launching the place to actually use the best ones:
decisionmodels.io
➡️Top open models
➡️Up to 4x faster on our own REAX engine
➡️From $0.019 per 1M input tokens.
Expect more reveals soon, and read 👇
Fezicles.eth retweeted
If you're concerned about the consolidation of control of AI by a few powerful companies/governments the only alternative is models so small anyone can run one on their laptop, phone or consumer hardware (e.g. like a DGX Spark). This decentralized vision is the future @jon_durbin has set @chutes_ai to building.
Most AI runs in a datacenter, and someone else holds the off switch.
At Exploit Summit (@ExploitSummit ), Jon Durbin showed Parallax running on a phone. CPU only, no GPU. The trained 8B model decoded at 59.6 tokens per second. A 40.75B model ran at 26.9 tokens per second with 12 GB peak process memory.
Where that leads:
• You could use AI in a country that blocks it.
• A doctor could run a model on a tablet in the exam room, and no patient data leaves the building.
• A law firm could keep one on a laptop in a locked office and never worry about prompts leaking to a vendor.
• You could use it on a plane or through an outage.
• No monthly bill and no rate limits. No company deciding tomorrow that your use case violates its policy.
A datacenter has an off switch. A billion personal devices don't.
Fezicles.eth retweeted
The push for regulatory capture and AI cabalism is accelerating. This, in large part, is why I'm building parallax. True P2P unstoppable, uncensorable AI systems aren't just neat, they're mandatory.
This little single DGX spark "datacenter" in my forge with parallax can train (with other nodes) an 80b parameter model.
- two solar blankets
- ecoflow battery
- lapdock for kvm
- starlink internet
- dgx spark
Total input from solar ~600w (800w max)
Total power draw from spark/starlink/etc. < 200w
No datacenters, no water supplies, no external power grid, no multi gigabit link.
We must design AI systems that are as resilient as the humble cockroach.
Why do we all appreciate BTC? Because it's a mathematically certain sovereign ledger that can't be inflated away or centrally controlled.
Current AI represent the greatest concentration of wealth, power, and information (IP theft? blackmail? or simply ads and behavior control?) the world has ever seen... As BTC did to finances, we must do to AI.
Chutes + parallax is a step forward towards tearing down this moat.
Fezicles.eth retweeted
Now we accelerate.
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.
You can read the full post here: darioamodei.com/post/we-must…
Fezicles.eth retweeted
DeepSeek V4 Flash runs on Chutes at $0.44 per million input tokens, $1.32 out. 1M-token context.
Inside a hardware-attested TEE with end-to-end encryption, like every model we serve. Nobody reads your prompts on the way through.
Try here: chutes.ai/app/chute/chutes-d…
your AI runs on infrastructure you've never seen, operated by people you'll never meet, inside servers you have zero claim to.
that's not a conspiracy take. that's the architecture.
three companies sit between your product and the compute it runs on. they set the price. they set the limits. they can cut you off.
Chutes runs on Bittensor.... decentralized, open-source, TEE-secured.
you can see the infrastructure. you can audit it. you're not renting a black box.
what else could you possibly want? 👇 chutes.ai
Do yourself a favor and add this to your Claude
Thanks
i extracted all the lessons from my 14,000 tweets & turned them into a browsable knowledge-base...
you can read them here: eptwts.com
Fezicles.eth retweeted
A 5B model is beating a 7B MoE trained on the same amount of data.
Early result from Jon Durbin's Parallax training runs. The bigger model has 30 percent more active parameters. Active parameters are the part of a model that actually runs when it answers you, so they are what compute bills are made of.
A win here compounds. Same quality for less compute, on every request the model ever serves.
Four techniques stacked make it work: quantile balancing, LatentMoE, a GDN-2/SWAX/MSA hybrid, role-differentiated precision. Parallax is betting the stack on exactly this kind of efficiency.
The Parallax paper: chutes.ai/parallax.pdf
Jon's full breakdown: nitter.gq/jon_durbin/status/2082…
Napkin math on bits per byte vocab adjusted etc, this is also beating olmoe 7b 1ba at matched tokens (which is 30% more active params, 40% more total params). Maybe competitive with MobileMoE if we did a complete several trillion token severely overtrained run.
Quantile balancing
LatentMoE
Hybrid GDN-2/SWAX/MSA
Role differentiated mixed precision
I think we have a winner. More napkin math, I think this 5b model would be able to do around 1200tps inference on 5090s.
Kinda want to do the full 9t token run on some 5090s of this tiny one, for funsies. Parallax-5b-instant (or we could bump total params up to 10b for free if we don't change active params, or maybe higher)
Interested to know how Parallax works?
Watch it here 👇
.@jon_durbin pre-trained a 20B MoE for under $10 an hour of compute.
DropZone Episode 1 is him explaining how, in full.
- a surrogate 48x smaller than the real expert, standing in for it
- workers that train on data they never see
- what ternary weights change about the hardware you need
- why the existing sync algorithms all failed, so he wrote one
Here's the full episode 👇
$TAO
Distillation only gets you as far as the teacher. You inherit the teacher's blind spots, and you cannot verify what it taught you.
Parallax does not distill. We build a graph database of real, grounded sources and derive each question from its answer, verifying every step.
The result is training data you can trust, not a copy of someone else's model.
$TAO
AI servers will burn 175 TWh of electricity in 2026, up from 95 TWh last year. That is Gartner's forecast, and it is nearly double in a single year.
The industry's answer is more gigawatts. Ours is fewer.
Parallax trained a 20B model on cheap, mismatched GPUs to within about 2% of a datacenter baseline. Efficiency is not a nice-to-have. It is survival.
Fezicles.eth retweeted
The goal was never another LLM. It was never even "an LLM that happens to train decentralized."
The goal is a sovereign, open-source platform for the whole model lifecycle: build it, serve it, fine-tune it, own it. Always available, to everyone, forever, borderless.
That is what Parallax and Chutes are for.
Fezicles.eth retweeted
Inference is it. It's the whole game. At the end of the day, any fine-tuning, RL, pretraining, prediction models, whatever else you're trying to build, if you don't use it, it's useless. Tooling almost exclusively focuses on test-time compute now vs. better base models.
And, at the same time, inference needs to be more efficient. DFlash/DSpark are great, but we need algorithmic/architectural breakthroughs, not just extra add-ons post-hoc.
Creating/curating data pipelines and synthetic augmentation/etc. also need inference as a prerequisite to train the model. Using agents = inference. RL = largely inference. The goal of training a model = inference.
That's parallax's "why" - we need to address the compute/cost/energy efficiency crisis facing the AI space right now. Maximally compact and efficient models to extract every last IQ point per watt out of the hardware, vs. just throwing more hardware at the problem.
A̶t̶t̶e̶n̶t̶i̶o̶n̶ inference is all you need.
What happens when the compute behind AI stops living in one company's basement?
Last month @chutes_ai served 1.23 Trillion Tokens. No data center. No single owner.
Open models running on GPUs worldwide, live on Bittensor.
A few giants want you renting AI forever. We went the other way.
Would you plug in? 👇
Fezicles.eth retweeted
How do you price an AI app when your heaviest user costs you 4x what they pay?
Flat $20/mo. Most users cost you $1 in inference. The power user burns $80 and you eat it. The more they love the product, the more they cost.
Sign in with Chutes flips it. Users bring their own Chutes account and credits. The tokens come out of their balance, not your margin. Cost per user trends toward zero.
Won't fit every product. But if your unit economics get worse with every active user, it's worth a look.
What's your AI cost per user?
You don’t own any of the AI you use.
You’re renting it from ~3 companies who set the price, the rules, and what you’re even allowed to build.
There’s another way:
Decentralised compute.
Thousands of people contribute GPUs on one open network.
Open models, secure inference and cheaper costs
All enabled by @chutes_ai
Starting to dig into the parts of AI nobody puts on a billboard.
This one’s the power problem.
Kicking it off over on @chutes_ai.
A few more on the way.
xAI built an AI training center the local power grid couldn't feed.
So the company brought in dozens of its own gas turbines to cover the gap, hundreds of megawatts, some running before they were permitted.
That's the real barrier to building a frontier model right now. Not a breakthrough idea. A private power plant and room for half a million GPUs.
Should building AI come down to who can build a power station? This is the lock we're working on 👀
