Razuki.eth

[Rohan Paul @rohanpaul_ai]

New Stanford paper argues that LLMs already provide the raw skills for AGI, but they still need a coordination layer on top.

Here the LLM is a fast pattern store, while a slower controller should choose which patterns to use, enforce constraints, and keep track of state.

To describe this, the author defines an anchoring strength score that grows when evidence clearly supports an answer, stays stable under small prompt changes, and avoids bloated, noisy context.

When anchoring is weak the model mostly parrots generic patterns and hallucinates, but past a threshold it switches into more reliable, goal directed reasoning, as small arithmetic and concept learning tests show.

MACI then runs several LLM agents in debate, tunes how stubborn they are from anchoring feedback, inserts a judge to block weak arguments, and uses memory to track and revise decisions on longer tasks.

The main claim is that most LLM failures come from missing anchoring, oversight, and memory instead of a bad pattern substrate, so progress should focus on building this coordination layer rather than discarding LLMs.

----

Paper Link – arxiv. org/abs/2512.05765

Paper Title: "The Missing Layer of AGI: From Pattern Alchemy to Coordination Physics"

[Mark Jeffrey @markjeffrey]

I am very excited to announce the launch of Stillcore Capital, a new Bittensor fund.

My partners are @jason and @rob_svrn.

Our focus will be $TAO and Bittensor subnet alpha tokens.

More at @stillcorecap and https://t.co/9jVPimJXpC.

[Marc Zeller @Marczeller]

Published one of my most important posts on the Aave governance forum.

If you want to know what the next steps and strategy we are supporting for Aave, have a read.

https://t.co/slJdMOxkeD

[Google DeepMind @GoogleDeepMind]

Introducing AlphaGenome: an AI model to help scientists better understand our DNA – the instruction manual for life 🧬

Researchers can now quickly predict what impact genetic changes could have - helping to generate new hypotheses and drive biological discoveries. ↓

[kache @yacineMTB]

I got fired today. I'm not sure why, I personally don't think there is a reason, or that it's important.

When I joined twitter, I joined because of the engineers I met in SF. They seemed happy. They were having fun. Engineers at play. Engineers that were enabled. It was good!

[Razuki.eth @Razuki_eth]

@ylecun Reduced focus

[Andy ττ @bittingthembits]

No one is ready for $TAO 💥

Around the 17 to 18-minute mark, macro investor Jordi Visser drops a bomb:

“More people are asking me about Bittensor than any other crypto, and they’ve never even bought crypto before.”

Family offices, TradFi, and AI-native capital are waking up. 👁️

Bittensor is about to become the gateway to AI yield, and $TAO is the only key.

🔗 Watch the clip:

https://t.co/fK608FwmXk

🎙️ @APompliano @jvisserlabs #Bitcoin #TAO #Bittensor #AI #Crypto

[Proof of Talk @proofoftalk]

The dream 𝜏eam is coming to Proof of Talk!

For the first time, Proof of Talk will host a dedicated Bittensor track, bringing together the leading minds behind the most exciting decentralized AI protocol in Web3.

We’re thrilled to unveil the official Bittensor track agenda THREAD.

If you believe in the future of decentralized AI, this is the one track you can’t afford to miss.

Book your ticket now and join the minds shaping the most advanced network in Web3.

Use code TAO for 24% off, only valid for the next 24 Hours!

🔗Secure your spot now: https://t.co/PjQJDqFGEN

[siamkidd @SiamKidd]

Had a lot of DMs over the weekend about a lot of chatter about the 'lack of dTAO uptake', root APY being too high and general fear about whether subnet tokens will ever recover (lol).

So here are my personal thoughts on it all:

1.) 'Lack of dTAO uptake'. Crypto peeps are always time horizon myopic. And that's because we've been spoiled by previous Crypto cycles. "We want 100Xs and we want it now darn it!" But Bittensor ISN'T a crypto project folks. 'We aren't in Crypto land anymore Toto'. It's the TCP/IP of decentralised AI and it just so happens that the only way for Yuma to properly work is to utilise a DLT. That's it. (Bit shit that Substrate is the DLT we've chosen but heyho, that's another topic altogether)! We just happen to have attracted a shed load of Crypto punters which is where most of the disappointed noise is coming from.

Bittensor is a brand new island that's popped up in the ocean and OTF/subnets are busy laying in hardcore infrastructure to this island. The equivalent of pipes, sewage, roads, bridges, energy etc etc. dTAO is something the world has never really seen before and it's almost 4 months old...chill out!

2.) 'Root APY being too high'. I sort of agree with this. When you analyse the top 100 TAO wallets, you'll see that basically 95% of them are in subnets. Just chilling in root. Why wouldn't they? Const constantly warned us all that the first 3 months would be chaotic as the dust settled and the APY has been great for them. I am a top 100 wallet holder and I know many others too. So from my perspective it's NOT just the APY that's holding whales back. It's the thin liquidity AND APY.

If you have a 50 000 TAO ($20m wallet), if TAO weights were changed and it dropped root APY down from 20% all the way down to 5%, would they move to subnets? I'd wager that the answer is NO. Even at 5% APY, they're still receiving 208 TAO per month (~$85k p/m) and even TAO at a measly $3000 that's $624k per month! Relatively risk free. Why risk losing TAO on subnets when there's generational wealth incoming from doing nothing? I don't blame them in hindsight. Depending on the wallet, I'm 70-100% in subnets and I see daily fluctuations of 1000ish TAO :-S In recent days, it's down only lol.

So in order to get the whales out of root, you need a really strong forcing function to prod them out of comfort. I don't know what that may be, but I guess 1-3% root APY may move a chunk of whales into dTAO, but not all...

3.) Liquidity! Everyone has this back to front! Prices aren't going down because liquidity is increasing! You need to understand that it's NOT retail that sustainably move prices. Yes, in teenie low liq markets (prevalently seen in NFTs and memes) retail can spike prices. But what happens after spikes? Yup, typically, big re-traces. So a LOT of the price surges in subnets thus far have only been because the 2000-3000 odd subnet investors have stampeded into very low liquidity instruments (subnets)! Now 3000ish subnet investors/traders are in, we've been in a capital circulating period over the last month or so where it's just the same bunch of TAO flowing in between subnets. Bittensor on X is a big echo chamber and noisey noise/announcements trigger that same capital to rotate in and out of subnets.

Other than root APY being cushy, the other main reason that many whales aren't participating in dTAO is because there is bugger all liquidity in the subnet liq pools! And it's whale participation that will lead to sustainable higher prices.

Bigger wallets can't even dip their toes into a subnet without the price spiking. Which means they can't get a proper position size in and therefore don't bother. Put yourself in their boots. They see a promising subnet, but because the LP is so shallow, they can barely even get 0.1% of their portfolio into it. So there's no point. And they're not going to bother buying tid-bits of alpha 5 times a day for 14 days straight just to accumulate a 1% portfolio position!

I do that, because I'm an obsessive geek with no life. But others won't. This is literally one of the reasons I'm actively hunting to do as many OTC deals as possible!

Also note that the top 15 subnets are about 80% of the sum of all subnets. That's why it's those that fluctuate the most when sum of all goes up or down. Why? Because that's where the liquidity is!

Deeper liq pools = more whale participation = higher SUSTAINABLE alpha prices!

Forget retail folks! The future success of Bittensor is NOT retail crypto adoption! It's hardcore big capital that will flow into this space once the subnets start shipping game-changing services or game changing models... There will also be a constant flow of big capital entering Bittensor. I'm having loads of chats with other funds and biz owners who are setting up fund raising vehicles to plough into TAO. Patience is key...

But for the retail lot, there is still a promising near future... Each subnet gets 1 TAO put into their LP every 12 seconds. In about 3 months time, most pre-dTAO subnets will have ample LP depth for more capital to easily flow into. So I'm still very positive that sum of all subnets should conservatively be around 3-4 by October time. And Reserves over Injected is screaming oversold at the moment. So we just need to wait for mean reversion to kick in! I'd be surprised if TAO isn't roaring a couple of weeks after Proof of Talk next week... That whole event is one big TAO pilling event for institutions and big money.

Bit of a ramble, hope this makes sense...what personally keeps my mind in focus is trying to maintain the bigger picture. For me it's this:

TCP/IP (Internet protocol) = A global network for information relay/dissemination.

Bitcoin protocol = A global network for information + wealth relay/dissemination.

Bittensor protocol = A global network for information + wealth + intelligence relay/dissemination.

Yup, Bittensor is as monumental as the creation of the Internet and Bitcoin. People just haven't figured this out yet...but if you're reading this post, you're so frickin early. As long as you don't shit the bed and dump your TAO, good times are practically coded in for you...

[const @const_reborn]

The largest LLMs will train under markets which align intelligence contribution.

Templar-I
https://t.co/y0gLDDU4Bx

[James Altucher @jaltucher]

The $TAO / @opentensor token is very interesting.

Question: why would any company build their own data center when they can build any AI model for 1/100 the price using the $TAO subnets?

With fixed supply of 21m tokens plus increasing demand for AI, this token seems like a no-brainer. Not to mention the staking possibilities using the subnets. Also, arb possibilities trading the subnets.

[Razuki.eth @Razuki_eth]

@kevinfinalbosu Letsugo!

[Igor Eckert @igoreckert]

Have you ever seen a meta-analysis of meta-analyses?

If it sounds statistically insane, you can rest assured it is.

But it's 2025, anything is publishable.

In the attached forest plot, each of these effect sizes is a different *meta-analysis*. All of them repetitively including many of the same trials.

This results in overlap. This overlap is called double-counting, and it's a terrible unit-of-analysis error.

Results are uninterpretable and artificially precise.

But they get published (and cited!).

[Andrew Zhao @_AndrewZhao]

Okay, I was definitely not vague posting

[Razuki.eth @Razuki_eth]

@neuromancer_t Ah yes... The paper alluded to it.

[Razuki.eth @Razuki_eth]

@neuromancer_t That's honestly a proof of concept at best. Error rates r still too high for fault tolerant quantum computation. And you can't really go w a decentralized structure right now to get qubits. Gated by well funded labs/orgs.

[LayerZero @LayerZero_Core]

LayerZero is live on Bittensor!

We’re excited to announce that LayerZero is powering interoperability for Bittensor and all its subnets.

With this integration, every ERC-20 asset on Bittensor's EVM can now expand natively to any of the 130+ chains LayerZero supports today.

The home of decentralized commodity markets, now connected to all of crypto.

[prayingforexits 🏴‍☠️ @mrexits]

Lee Kuan Yew edit goes insanely hard

[Razuki.eth @Razuki_eth]

@damskotrades $tao pls ser

[Andrej Karpathy @karpathy]

We're missing (at least one) major paradigm for LLM learning. Not sure what to call it, possibly it has a name - system prompt learning?

Pretraining is for knowledge.
Finetuning (SL/RL) is for habitual behavior.

Both of these involve a change in parameters but a lot of human learning feels more like a change in system prompt. You encounter a problem, figure something out, then "remember" something in fairly explicit terms for the next time. E.g. "It seems when I encounter this and that kind of a problem, I should try this and that kind of an approach/solution". It feels more like taking notes for yourself, i.e. something like the "Memory" feature but not to store per-user random facts, but general/global problem solving knowledge and strategies. LLMs are quite literally like the guy in Memento, except we haven't given them their scratchpad yet. Note that this paradigm is also significantly more powerful and data efficient because a knowledge-guided "review" stage is a significantly higher dimensional feedback channel than a reward scaler.

I was prompted to jot down this shower of thoughts after reading through Claude's system prompt, which currently seems to be around 17,000 words, specifying not just basic behavior style/preferences (e.g. refuse various requests related to song lyrics) but also a large amount of general problem solving strategies, e.g.:

"If Claude is asked to count words, letters, and characters, it thinks step by step before answering the person. It explicitly counts the words, letters, or characters by assigning a number to each. It only answers the person once it has performed this explicit counting step."

This is to help Claude solve 'r' in strawberry etc. Imo this is not the kind of problem solving knowledge that should be baked into weights via Reinforcement Learning, or least not immediately/exclusively. And it certainly shouldn't come from human engineers writing system prompts by hand. It should come from System Prompt learning, which resembles RL in the setup, with the exception of the learning algorithm (edits vs gradient descent). A large section of the LLM system prompt could be written via system prompt learning, it would look a bit like the LLM writing a book for itself on how to solve problems. If this works it would be a new/powerful learning paradigm. With a lot of details left to figure out (how do the edits work? can/should you learn the edit system? how do you gradually move knowledge from the explicit system text to habitual weights, as humans seem to do? etc.).