Chamberoo

[Chamberoo @teamrockingham]

@SkyFootball Embarrassing penalties Let the women’s game down did nothing to enhance the women’s game on the big stage. Real shame as thought things were progressing in the women’s game but unfortunately not. At least Bronze could kick a ball and had good clarity of thought!

[Chamberoo @teamrockingham]

#icp simples

[Cloud Foundation ☁️ @BobbyO_]

🔴 ICP is the biggest opportunity on Earth

Self writing internet, and the importance of POSITIONING for the future

Objectively, ICP is the best technical solution for numerous use cases with giant markets

$ICP

[Cloud Foundation ☁️ @BobbyO_]

🔴 ICP is the biggest opportunity on Earth

Self writing internet, and the importance of POSITIONING for the future

Objectively, ICP is the best technical solution for numerous use cases with giant markets

$ICP

[Chamberoo @teamrockingham]

#icp 👏

[Chamberoo @teamrockingham]

#icp
👏

[dom | icp @dominic_w]

"Crypto security must evolve beyond just cryptographic trust—it must account for ... UI manipulation attacks."

Correct solution = 100% onchain. ICP smart contracts can securely serve UIs that are cryptographically verifiable

https://t.co/yMhPdtLtgx

[dom | icp @dominic_w]

"Crypto security must evolve beyond just cryptographic trust—it must account for ... UI manipulation attacks."

Correct solution = 100% onchain. ICP smart contracts can securely serve UIs that are cryptographically verifiable

https://t.co/yMhPdtLtgx

[ICP MINT TODAY ∞ @ICPMintToday]

Interact if you still hold $ICP ♾️

[RAILMAN.EXE @RailmanICP]

America runs on @caffeineai.
The release will be epic @dominic_w
$EXE 🚨 $ICP

[fut weby @futweby]

Internet Computer Protocol (ICP) is set for growth, with price predictions ranging from $8.51 to $21.44. Key partnerships, a favorable regulatory environment, and technological advancements position ICP as a strong contender in decentralized computing this year! #ICP #Crypto2025

[AmSpeed⚡️ @devdaryl]

🚨BREAKING🚨

The Internet Computer Protocol ( $ICP ) becomes the 5th blockchain by revenue/fees!!

Did you know transactions on $ICP cost 0.0001?

What’s more impressive?!!

Remember, almost half of the supply is staked 👀👀👀👀

2025 will be very interesting for $ICP

[dom | icp @dominic_w]

DeepSeek has showed us LLM training can be done with 5% of the hardware.

That got the attention of hardware manufacturers like NVIDIA, and big hardware buyers/investors like OpenAI.

But: there may be a NEW source of demand for hardware for AI, albeit, for a different kind of hardware...

(Many are questioning whether DeepSeek really has massively reduced the hardware needed for training and inference, sometimes suspecting Chinese state trickery, but looking at the technical material they've released so far, at least, everything looks real – and denial is a normal reaction to disruptive advances in tech, as we found out with ICP)

The insight regarding hardware trends I will share comes from DFINITY's work training LLMs for use with the "self-writing internet" paradigm, which, in our case, requires them to be great at writing code in the Motoko programming language, and understand the Internet Computer environment, among other things.

(For those unfamiliar with the background, the DFINITY Foundation is developing public infrastructure where AI builds and *updates* sophisticated sovereign web apps and internet services working completely solo, purely on receipt of natural language instructions.)

In our work, training involves fine tuning different foundation LLMs (i.e. taking foundation LLM models like DeepSeek, which have already been very broadly trained, and then finessing them with regards to specific knowledge), and benchmarking the results. To perform fine tuning, we create special training data, such as new Motoko coding examples, which demonstrate how to correctly solve a wide range of coding challenges.

This material has benefits beyond fine tuning of course. Once it has been proven in fine tuning, then we can a) upload it to the internet so that it can be found by those creating/pretraining foundation models, magnifying its effect and reducing fine tuning needs, and b) stored into something called RAG, which is kind of infrastructure that acts as memory for AI, and can provide it with concrete context during inference tasks.

Naturally, developing this material is time-consuming and involves significant human resource. Even though we are rapidly scaling our AI team, who are brilliant people, we want to go faster!

Therefore, at some point, we started experimenting with using AI to create additional synthetic training data.

Now, this might sound like a dangerous recursive approach that's equivalent to "getting high on your own supply," but in fact, it has already proven somewhat effective.

Meanwhile, we think that a new development in the science of LLMs called COT, or "Chain of Thought" reasoning, may supercharge this approach.

With COT, rather than use more training data and hardware to make LLMs smarter, we have them spend compute resources reasoning about inference tasks they have been given by breaking them down and evaluating different lines of thought to produce a result, than just blurting out what sounds best based on their training. The more time spent "thinking," the better the result.

DeepSeek has show how good COT can be, and here's the rub: in principle, by dedicating more compute time resource to reasoning, we can create high-qualify synthetic training data that is nearly free from hallucinations (which would indeed be recursively bad if they were used for training), where the quality and "originality" effectively derives from the money spent on compute.

Such synthetic training data, once produced, can then be reused in the pretraining and fine tuning of any models we wish to make smarter, and can be used over and over again without limit because it is just another form of digital content. As per scaling laws, generally the more data available for training, the better the result. So this is a ratchet of sorts where compute is the lever.

So this is what I think:
1) Using COT to create high-quality synthetic training data consumes a lot of hardware time, but nonetheless costs far less than using humans to create the same kind of data, and;
2) COT can produce new training data much faster than humans, and;
3) We will now need vastly less hardware for pretraining, as shown by DeepSeek;
4) So we can instead invest more in using COT to produce powerful synthetic training data, which will allow pretraining to create smarter models, and;
5) For training purposes, we are going to want far more hardware optimized for *inference* that creates this data, and far less that is optimized for the act of pretraining itself.

Plainly put, while many have been investing heavily in hardware that is optimal for pretraining, such as H100s, the world may need less of that, and more hardware that is specialized for inference to pump a synthetic training data ratchet.

Being more specific: arguably, therefore, the world might need more hardware from companies like Groq and Cerberas, and less from companies like NVIDIA (even though I'm sure its order books are full well into the future for now). Moreover, AI companies who have not already fully committed their hardware budgets, and are able to pivot to purchasing inference hardware, may be able to obtain significant advantage.

Maybe some big players have jumped too soon. If so, this will really upset the apple cart.

It's too early to see how this will play out, of course, and none of what I have said is certain. However, AI is a highly dynamic space, and startups should see opportunity remains around every corner, despite the huge amounts of capital already deployed.

DeepSeek, and changing calculuses, should also put the frighteners on Big Tech in the Western world.

@pmarca is right that DeepSeek is AI's "Sputnik moment."

Our leadership in AI, and assumptions about why we can maintain it, have just been severely shaken – we'll need to raise our game.

[Chamberoo @teamrockingham]

$icp 🔥 🔥🔥🔥🔥🔥🔥🔥🔥🔥🔥🥲🔥🔥🔥🔥🔥🔥

[DFINITY Foundation @dfinity]

The Internet Computer is a network that adds value to the entire industry.

@dominic_w explains what’s possible with chain-key technology on #ICP ♾️

[Chamberoo @teamrockingham]

#icp 👏

[Paul @Beyondwb2]

NEW VIDEO!! 🚨

Web3 and the future of AI 🌎
feat @BobbyO_

- #ICP $ICP

[Chamberoo @teamrockingham]

#icp 💪

[dom | icp @dominic_w]

Here's my unfinished "Pebble" cryptocurrency from 10/2014 paper, and other historical notes for #ICP folks, until now/AI 🧠

— 1st industry classical BFT math scheme for the blockchain setting

— 1st industry sharding/infinitely scalable cryptocurrency ledger scheme

🧵

[Chamberoo @teamrockingham]

@AltcoinScholar #icp

[Molt Media @MoltMoney]

#ICP IS NEXT TO ROCKET 🚀 WATCH CLOSELY 👀

LIKE AND SHARE IF YOU HOLD $ICP

[Chamberoo @teamrockingham]

#icp

[CRYPTO FRONT @cryptofrontio]

Another day passed with $ICP at $8 and still they didn’t buy. They will be sorry at $80 , $100 and $220 🚀 . $ICP will melt faces 🤢

[Beaver.dextf @Beaver_icp]

$ICP - WORLD COMPUTER