Lex Tensor (@Lex_Tensor) profile picture

Fractional GC and τ since 2022. None of my posts are legal advice.

The Neural Network
Joined June 2026
Photo by Lex Tensor (@Lex_Tensor)Photo by Lex Tensor (@Lex_Tensor)Photo by Lex Tensor (@Lex_Tensor)Photo by Lex Tensor (@Lex_Tensor)Photo by Lex Tensor (@Lex_Tensor)Photo by Lex Tensor (@Lex_Tensor)Photo by Lex Tensor (@Lex_Tensor)Photo by Lex Tensor (@Lex_Tensor)Photo by Lex Tensor (@Lex_Tensor)Photo by Lex Tensor (@Lex_Tensor)
Very bullish for @chutes_ai (and Bittensor) who are building a solution to this issue bittensor:native
This might be the end of this GPU cycle... The credit bubble will pop, the data centers will not be built It will give us 2-3 years to get thermodynamic computing to trillion parameter scale We will then begin the next scaling era with power-efficient compute Trust the plan.
15
Lex Tensor retweeted
and btw, we can't do any of this stuff properly in Web2. Crypto will never die.
OSS AI is just a catch all. What we need is permissionless compute, non KYC inference, encrypted model storage, decentralized pretraining, and collaborative RL. Super intelligence is either co-created and co-owned with free access or it will become cancerous and asymbiotic with humanity.
22
34
1
295
9,436
Bittensor subnet @chutes_ai has partnered with Tencent Cloud. Chutes is a launch partner for @TencentHunyuan’s new flagship AI model, Hy4, now available through its API. Day-one access to leading AI models is business as usual on Bittensor.
Tencent Hy4 is live on Chutes We’re a launch partner with @TencentHunyuan. It’s Tencent’s new flagship, and the first Hy model that can see. It reads screenshots and charts, and it can drive a terminal or a desktop app for you. 0.525$ per million tokens in, 1.5625$ out. Already on Chutes? Just Change the model name to start using it Try now: chutes.ai/app/chute/chutes-t…
14
56
6
332
14,353
Lex Tensor retweeted
Hy4 is coming to Chutes 🥳
Tencent Hy4 is coming to Chutes 🪂 We're partnering with Tencent Cloud to bring it to the same Chutes API you already use. Same key, new model ID More soon 👀
3
89
6,719
Lex Tensor retweeted
next two weeks: sillicon valley but first: sancerre valley
3
3
1
91
2,231
Lex Tensor retweeted
October on SN44, in a few lines. -More public tracking tasks running on the subnet -3 new public elements out -A big chunk of example data from the cricket private track is now public -A lot of quiet work on security and scale -We can spin up new tasks faster now, with plenty more coming And following IBC, one of the largest real-time graphics software companies is now testing a Studio endpoint, running on the football model built earlier this year under our previous incentive mechanism. Meanwhile, we’re collecting feedback from Studio users for what comes next. From here the pace only picks up, with way more tasks every month and one goal, making computer vision a commodity.
8
15
3
111
4,026
Lex Tensor retweeted
Chutes is very excited to be working with METANOVA LABS as they develop their Hit-to-Lead optimization agent! We fully intend to assist and support to this project in any way we can as it develops.
NOVA Hit-to-Lead Optimization Agent In collaboration with @chutes_ai and @onepot_ai We're developing a Hit-to-Lead optimization agent that integrates an algorithm developed by a NOVA Blueprint winner, extending NOVA's capabilities beyond hit discovery and into the iterative optimization of small molecules. The system combines a domain-specialized AI agent with proprietary combinatorial chemical databases supplied by our partner Onepot AI, spanning billions of possible compounds, that are delivered within 5-7 business days. Researchers will be able to guide the agent in developing, comparing and refining optimization strategies, exploring chemical modifications and prioritizing candidates based on predicted affinity, selectivity, physicochemical properties and specific binding interactions. Confidentiality is a central design requirement. Pharmaceutical R&D involves valuable intellectual property, including compound structures, targets, emerging chemical series and optimization strategies that cannot be exposed publicly. The Chutes team integrated domain specific models so that we can run inference in their verified Trusted Execution Environments (TEEs), with encrypted communication and controlled access to proprietary databases and project records. This architecture is intended to support confidential inference on external infrastructure while preserving organizational control over sensitive research data. Finding a molecule with promising predicted affinity is only the beginning. Hit-to-Lead optimization involves balancing multiple, often competing objectives, including potency, selectivity, physicochemical properties and synthesizability, before progressing toward experimental activity and ADME validation. The next step is to integrate the agent into our own Hit-to-Lead development program. For NOVA, this represents an important step toward translating algorithms developed through open competition into tools for proprietary drug discovery programs, including our own asset-development pipeline and those of external research partners.
3
21
112
2,687
NOVA Hit-to-Lead Optimization Agent In collaboration with @chutes_ai and @onepot_ai We're developing a Hit-to-Lead optimization agent that integrates an algorithm developed by a NOVA Blueprint winner, extending NOVA's capabilities beyond hit discovery and into the iterative optimization of small molecules. The system combines a domain-specialized AI agent with proprietary combinatorial chemical databases supplied by our partner Onepot AI, spanning billions of possible compounds, that are delivered within 5-7 business days. Researchers will be able to guide the agent in developing, comparing and refining optimization strategies, exploring chemical modifications and prioritizing candidates based on predicted affinity, selectivity, physicochemical properties and specific binding interactions. Confidentiality is a central design requirement. Pharmaceutical R&D involves valuable intellectual property, including compound structures, targets, emerging chemical series and optimization strategies that cannot be exposed publicly. The Chutes team integrated domain specific models so that we can run inference in their verified Trusted Execution Environments (TEEs), with encrypted communication and controlled access to proprietary databases and project records. This architecture is intended to support confidential inference on external infrastructure while preserving organizational control over sensitive research data. Finding a molecule with promising predicted affinity is only the beginning. Hit-to-Lead optimization involves balancing multiple, often competing objectives, including potency, selectivity, physicochemical properties and synthesizability, before progressing toward experimental activity and ADME validation. The next step is to integrate the agent into our own Hit-to-Lead development program. For NOVA, this represents an important step toward translating algorithms developed through open competition into tools for proprietary drug discovery programs, including our own asset-development pipeline and those of external research partners.
4
15
5
67
5,827
Lex Tensor retweeted
Confidential open-weight models from @chutes_ai sit right next to claude, gpt & gemini on SayGM. Same key, same code. Your prompt stays sealed twice: once in our gateway TEE, again in chutes’ attested hardware. SN28 x SN64 🤝 saygm.com/blog/access-chutes…
15
2
102
1,715
Lex Tensor retweeted
First time I've seen a roadmap target anywhere for a "non-extractable" artifact in the wild, despite my suspicions about others working on similar. This is an exceedingly interesting development. The upside to such products is potentially an industry game changer, though as of yet unproven and only explored in pioneering whitepapers such as Pluralis. I suspect it will eventually be a very common, highly variable primitive and will attract very large sums from investors and intense loyalty and development from holders. I believe, as with many things at the intersection of crypto and AI, that $TAO will be among the best places, if not the best place, to launch such products, and that certain subnets are uniquely suited to take advantage of this line of thinking. $TAO SN64
Here's a very crude simplified chutes roadmap/plan, for near term: - move the most unprofitable things off of the subnet, including today's announcement of some (not all) additional inference options (not paid via emissions) - this allows us to scale the things we want and need to avoid capacity issues/better direct incentive - this includes models we use to build data pipelines and relationships to other teams/networks/etc - add options that are completely revenue positive (hourly pricing/rentals/simple docker based jobs/etc), TEE, private, secure, priced above cost, more extensive version of private chutes we have today basically - expand the hourly compute offerings to also include CPU only (e.g. to run agents, web browsers, sandboxes, non-GPU work of any kind) - use some of the incentive for ideally non-datacenter based model training and serving - build a non-extractable product on chutes from the models we build, via dynamic edge nodes, possibly from the epigenesis work (smaller faster cheaper faster than lora customizations from the sparse sections of pretrained model, participant supplied) - models and pipeline etc. will be open source, the customization and ensemble thereof will live only in chutes, inc. the continuously evolving set of served variations/RL/etc. We'd like, to the extent we can, attract participants who want to build and have no obligation to pay datacenters or any particular hardware vendor, to keep the value inside the subnet at all times by attacking sell pressure at the source (DC commitments, chip purchases, etc). We will continue to have secure TEE inference and GPUs; not elimating anything wholesale, just focusing more on a commodity that doesn't have to flow out to DCs and single hardware vendors. NVIDIA reportedly has 75% profit margins, revenue for the sake of revenue fueling GPU reselling goes straight to one place, basically. Fine, I guess, but I'd rather work together as individuals with whatever hardware we can bring, to a reasonable degree.
1
1
6
554
Lex Tensor retweeted
Here's a very crude simplified chutes roadmap/plan, for near term: - move the most unprofitable things off of the subnet, including today's announcement of some (not all) additional inference options (not paid via emissions) - this allows us to scale the things we want and need to avoid capacity issues/better direct incentive - this includes models we use to build data pipelines and relationships to other teams/networks/etc - add options that are completely revenue positive (hourly pricing/rentals/simple docker based jobs/etc), TEE, private, secure, priced above cost, more extensive version of private chutes we have today basically - expand the hourly compute offerings to also include CPU only (e.g. to run agents, web browsers, sandboxes, non-GPU work of any kind) - use some of the incentive for ideally non-datacenter based model training and serving - build a non-extractable product on chutes from the models we build, via dynamic edge nodes, possibly from the epigenesis work (smaller faster cheaper faster than lora customizations from the sparse sections of pretrained model, participant supplied) - models and pipeline etc. will be open source, the customization and ensemble thereof will live only in chutes, inc. the continuously evolving set of served variations/RL/etc. We'd like, to the extent we can, attract participants who want to build and have no obligation to pay datacenters or any particular hardware vendor, to keep the value inside the subnet at all times by attacking sell pressure at the source (DC commitments, chip purchases, etc). We will continue to have secure TEE inference and GPUs; not elimating anything wholesale, just focusing more on a commodity that doesn't have to flow out to DCs and single hardware vendors. NVIDIA reportedly has 75% profit margins, revenue for the sake of revenue fueling GPU reselling goes straight to one place, basically. Fine, I guess, but I'd rather work together as individuals with whatever hardware we can bring, to a reasonable degree.
9
21
4
114
6,083
Lex Tensor retweeted
One of the things I’ve been testing at @webuildscore. Can’t say much yet, but you can see the difference. On top, the boxes jump around and miss half of what happens between frames. Below, they just stay locked on. It’s a small thing until you’ve watched the choppy version for a few weeks. More soon.
1
8
2
38
4,617
Lex Tensor retweeted
Intelligence has quickly become one of the most critical resources on earth, and unfortunately it has already turned into a race for unimaginable wealth and concentrated power. A small handful of companies own the frontier models, the data, and the hardware. We at Chutes have strong feelings about that. Read our thoughts on a different vision for the world and path forward here: chutes.ai/news/democratizing…
2
10
1
56
1,736
Lex Tensor retweeted
Replying to @Fiskantes
Bitcoin mines for SHA256 hashing power. Bittensor mines for all the elements of intelligence (be it compute, inference, pre-training, post-training/RL, synthetic data, applied AI, etc), and it tokenizes them. Pearl mines for MatMul hashing power. Suppose “Bitcoin of AI” is in the eye of the beholder.
Replying to @Evan_ss6
Ok, I just put together a proper effort post here, so hopefully you read it, but the TL;DR is that TAO hits all the notes you mentioned above. You want a coin that's essential to an AI business, where that business generates real-world value and that value drives back to the coin itself. That's Bittensor. I am going to assume you don't know much about Bittensor beyond the "AI Bitcoin" moniker and start with a brief origin story, since I think it's important and helps illuminate where "AI Bitcoin" came from. The interesting thing about Bitcoin for the Bittensor story is that it figured out how to use a token to bootstrap a decentralized compute network, owned by nobody and accessible to everyone, that to this day is the largest computing network on Earth. Bittensor's founders are AI researchers. They knew how important compute was going to be for building AI, and they set out to build the same thing for intelligence: a decentralized intelligence network, owned by nobody and accessible to everyone. Just as Bitcoin could coordinate a disparate group of people to fight centralized control of money, perhaps a new primitive, Bittensor, could be designed to fight the coming centralized control of intelligence. So they abstracted the Bitcoin incentive system (i.e., spend energy securing the network by solving SHA-256 hashes, get paid in BTC) so it could be applied to a whole new set of problems, all aimed at one goal: mining for intelligence. Mining SHA-256 today on Bitcoin is incredibly difficult, but as a system it's simple and objective. Mining for intelligence is orders of magnitude more complex. It's subjective, so it has to be scored on a gradient (study: Yuma Consensus), and there are multiple components that all need to be mined (compute, data, algorithms, etc.). That's where subnets come in. Each one tackles a piece of the problem. (1) Outside revenue + value to the tokens: It took years and a lot of iteration, but Bittensor is now doing exactly what it set out to do. It's mining intelligence. Compute/inference: ~5 subnets are coordinating GPUs around the world to serve raw compute and inference into the open market. Customers range from retail to enterprise, with the crypto abstracted away entirely. Estimates vary, but I put this group at ~$40-50MM aggregate ARR and scaling fast (for context, think ~200% QoQ and ~50% MoM growth). Training: two subnets are doing SOTA decentralized pre-training right now (SN3 and SN9). SN9 just wrapped up a 100B-parameter run with the highest MFU reported for a distributed pipeline-parallel run. SN3's current run is a 110B-parameter model, the largest decentralized pre-train ever attempted. These teams are past proof of concept and working on scale. And SN120 Affine is post-training Chinese OSS models at SOTA levels. Applied AI: a whole set of subnets incentivize miners to build verticalized models that power real products. Score for computer vision. RedTeam and Bitsec for cybersecurity. Minos for genomics. Yanez for identity and fraud. Trishool for alignment. Vidaio for video compression. I'm missing plenty more. These are real products being used in the wild by enterprises. I'd put this category at $10-20MM ARR, also scaling quickly. For context, 12 months ago I'd have said ARR across the entire network was under $5MM. The TAMs here are massive, and the blue-chip subnets are exiting the pre-revenue stage and entering the scaling-revenue stage right now. Critically, these teams are taking that revenue and buying/burning their tokens. Not because they're altruistic. Because the network makes it the only rational move. Two reasons for this: (1) A subnet pays its miners in its own token, a fixed number of tokens per block. So the only way to grow the supply side (more GPUs, better models, more output) is for the token to go up. If you're a compute subnet and want more capacity, you take your revenue, cover miner COGS plus some margin, and buy your token. Buy pressure on your token becomes a derivative of your compute bill. The token is the production budget. (2) And the chain runs the same game one level up. Subnets compete against one another in a fight for daily $TAO emissions, and the scoreboard is the price of each subnet's alpha token (a clever way to let the broader market instruct the protocol on how to allocate its resources). The winners get a bigger share, and for mature subnets that TAO goes straight into buying their own token off the pool. Fall too far in the ranks and you not only lose out on emissions, you risk being deregistered. So you might astutely point out that revenues < emissions today. That is true, but the trend is our friend. Bittensor is essentially running the Uber playbook, which I would argue a cryptocurrency is tailor-made to do. Use the token, rather than VC funding, to bootstrap supply before the demand exists. Once supply is there, demand scales and the flywheel kicks in: demand up → revenue up → buy+burns → alpha price up → miner rewards up → more/better supply → repeat. The early alpha buyers were the pre-revenue seed/Series A investors. Now we're entering the revenue/scaling era. Although revenue isn't yet covering emissions, as discussed above, it is growing. Meanwhile, the supply side is in its first halving cycle. TAO had its first halving last December, alpha halvings are coming in 2027/28, and the subsidy shrinks on schedule as revenue takes over. I am not speculating here either - a couple of subnets already bring in more from customers than they pay out to miners. Important to mention: to buy and burn its alpha token, a team has to buy TAO first and stake it into the pool that pairs TAO against its token. Every alpha burned leaves TAO behind in the pool. ~2M TAO already sits in those pools, and every subnet token can only trade against TAO. So subnet revenue = TAO buy pressure, mechanically. Speculative investing in subnets = TAO buy pressure. Subnet/miner/validator registrations = TAO buy pressure. Taken to its logical end... TAO is literally an index of all 128 subnets, or in other words, an index of 128 AI startups building on crypto rails. (2) It's cypherpunk: TAO also represents the right ethos for crypto x AI. It leverages a novel crypto primitive to bootstrap networks that coalesce edge compute and edge talent from around the world, coordinate them to produce the real inputs to intelligence (compute, data, algos, etc.), and serve them to anyone, anywhere. A network that serves open, permissionless, decentralized, uncensored intelligence. A true alternative to centralized labs and governments. And because it harnesses the world's edge compute and talent, it can scale in ways the fiat companies just can't. (3) You're still early: The top 3 compute subnets are ~$100MM market caps. The next two are $30-60MM. Promising subnets in other categories are sitting at $10-20MM. There are 128 subnets today and there'll be 256+ down the road. Maybe many more in the future. Bad teams get deregistered and new ones take their place. As the ecosystem grows, so do miner and owner rewards, which pulls in more teams and better talent. There will be fantastic teams launching subnets a year from now that you can buy at $5-10MM or less. CT isn't woke to this. 99% have not bridged over to the native Bittensor chain to participate in this ecosystem. Cross-chain bridges are only just going live. In my mind, this screams opportunity. And that's before you get to TAO itself, the L1/index/launchpad under all of it, sitting at ~$3B. That's comically low. Its 1:1 comp is ETH at ~$327B. 100x with size is viable here. Will follow up with some longer-form sources to check out.
2
5
518
So it's almost been 3 days since this post and Binance is apparently still out of $TAO. Reason is "high withdrawal volumes" - really nice to see a token listed by the cartel but where they don't hold that much of the supply (they had to list at their own initiative and not be handed several percentage of the supply as they usually get from projects listed there).
Barry was right, Binance is running out of supply bittensor:native
27
Barry was right, Binance is running out of supply bittensor:native
1
1
45
Lex Tensor retweeted
Yes. CPU only Chutes already finishing up testing. Current boot time is ~10 seconds, TDX and AMD support. Goes great with the models @jon_durbin has been working on for Parallax as well, looking extremely efficient for CPU inference. The preemtable issue and storage across instances is being worked on but probably a bit further out. I think solving this for a decentralized network will be a combination of engineering and incentive on the network.
1
2
6
110
Lex Tensor retweeted
Bittensor Subnet 64, Chutes, has an infernece platform that at one point was serving billions of tokens a day on OpenRouter. The problem? It wasn't profitable. In order for the subnet to survive long-term in the fierce competition that is Bittensor, this had to be figured out. The network incentivized Chutes to serve inference more efficiently. So backend dev @jon_durbin set out to solve it. In the process, he (and the Chutes team) may have wound up discovering an entirely new way to train models that could revolutionize decentralized training and truly unlock open-source models for the world. Highly recommend giving a listen to the story: youtube.com/watch?v=4FOxXT66…
1
7
57
3,547
Lex Tensor retweeted
Replying to @ExploitSummit
@ExploitSummit was an amazing event! So many brilliant, motivated people. If you want to hear about the model training stuff we've been working on, and why we're working on it in the first place, here's the talk. youtube.com/watch?v=4FOxXT66…
19
29
7
169
20,123