šŸ‘‘ Weekly AI Edge #9

Eric Schmidt (ex-Google CEO) gets spicy on AI

GM! Welcome to Chain of Thought, the best research on decentralised AI. Youā€™re in great hands.

In this edition of our weekly AI Edge, we cover:

  • Bittensorā€™s Subnet 11 Matches GPT 3.5 Turbo and Llama 3B

  • Sahara AI raises $43M from Pantera, Binance, and others

  • Eric Schmidtā€™s spicy takes on AI

  • Our favourite tweets on X

Elon might be enjoying Grok 2 a bit too muchā€¦

šŸ¦ State of the Market..

Source: CoinGecko

Over the past two weeks, the crypto x AI market (along with the rest of crypto) has taken a tumble.

The total market cap dropped 14% compared to 2 weeks ago but is up 7.7% over the last week.

TOP AI TOKENS

Price (7-day change)

FDV

NEAR

$4.09 (+8%)

$4.5 Billion

TAO

$268 (flat)

$5.6 Billion

RNDR

$4.64 (+4%)

$1.8 Billion

IO

$1.61 (-7%)

$1.3 Billion

ATH

$0.066 (+7%)

$2.8 Billion

 šŸ“Š Chart of the Week

Subnet 11 Matches GPT 3.5 Turbo and Llama3 8B

The top miners on Bittensorā€™s Subnet 11 (a roleplaying AI subnet) submitted models that outperformed big tech heavyweights like GPT-3.5 Turbo and Llama 3 8B in EQ Bench Scores.

EQ Bench measures the emotional intelligence of LLMs, and the results are telling. Llama 3 and GPT-3.5 Turbo scored 68.8 and 71.7, respectively, while models with a subnet score above 0.325 surpassed GPT-3.5 Turbo.

Two key takeaways:

1. Incentives work. Miners on Bittensor need to produce top-tier models to earn TAO tokens, and some are now competing with or even outperforming leading proprietary models.

2. Smaller models can excel in specific tasks. While GPT-3.5 is a massive model with over 100B parameters, the submitted models here are just 7-8B parametersā€”yet theyā€™re proving to be better at certain tasks.

šŸ† Caught Our Eyes..

Source: Sentient

  • Sahara AI announced a $43M raise from Pantera, Binance Labs, Polychain, and others. We explored Sahara AI in our recent research highlight.

  • Sentient reveals the Sentient Protocol, enabling AI models to be openly downloadable yet monetisable.

  • Nous Research unveils Hermes 3, a fine-tuned open model of Llama 3.1 that is uncensored and steerable

  • Peaq and Map Metrics starts pre-orders for their drive-to-earn car device

  • Sharpe AI, an AI-powered super app, announces their season 1 airdrop

  • Masa partners with Atoma to power AI inference with real-time data

  • Ora Protocol launches staking phase 2, staking with a new OLM and stablecoin pool

  • Spheron Network doubles its reward pool to $100k for builders on its decentralised GPU network

  • Kuzco Network announces a new chatbot powered by decentralized GPUs

  • SphereOne partners with Aizel Network to bring verifiability and privacy to its AI agent network

  • RoboNetā€™s points program for its AI DeFi platform goes live on Aug 13th

  • Lumerin Protocol, in partnership with 6079.ai and Morpheus, launched a Galxe campaign with $500 in prizes.

  • Yai partners with AtivoLabs to bring RWA tokenization through AssetLink.

  • Inference Labs partners with Symbiotic to bring cryptographic certainty to decentralized AI models

    Hyperbolic

  • Hyperbolic announces the alpha release of their GPU marketplace with free credits for selected testers

  • Hyperbolic makes Llama 3.1 405B Base available for researchers and developers on their network

šŸ§  Eric Schmidtā€™s Spicy AI Takes

If thereā€™s one video you should watch this week, itā€™s Eric Schmidt (ex-Google CEO) talking about AI to a class of Stanford students. Trust usā€”itā€™s 45 minutes well spent.

The video was posted on YouTube but was taken down. We found this version on X.

Just a few of his gems:

  • AI agents + text action + large context windows at scale will fundamentally change the world.

  • The leaders in propriety models are pulling ahead, and theyā€™re only going to get much, much better than the rest.

  • Nvidiaā€™s biggest moat? CUDA. Itā€™s like C++ for GPUs; many open-source libraries and tools that AI developers use are built on CUDA.

  • Tech giants are planning for increasingly large investments into Nvidia-based AI data centres, some >$300M

  • Current foundational models take about ~18 months to train, so theyā€™re always a bit outdated (6 months prep, 6 for training, 6 for fine-tuning). However, large context windows give AI more recency.

  • AI is a rich personā€™s game: Itā€™s all about capital and talent.

  • Thereā€™s a growing belief among investors that the invention of intelligence offers unlimited returns.

 

šŸ”„ On X..

Dither reviews the performance of their Pump Fun Bonding feed

Foresight Venturesā€™s AI x Crypto Report

Immortal Tofu on Zama fhEVM

Only Arweave with a breakdown of AstroUSD

Coovxrp.eth on how to run a Ar.io network gateway on a Raspberry Pi 4

DeFi Dad with a new Edge Pod

Delphi Digital unlocks Part 1 of their AI x Crypto report. Worth a read

Cheers,

Teng Yan & Joshua

This newsletter is intended solely for educational purposes and does not constitute financial advice. It is not an endorsement to buy or sell assets or make financial decisions. Always conduct your own research and exercise caution when making investment choices.

Reply

or to participate.