A few days ago @StanfordHAI just published the 456-page AI Index Report 2025. Here are some key takeaways for the AI trend:
🧵👇👇 1. First, AI becomes far more powerful than we think. AI performance in new benchmarks—MMMU, GPQA, and SWE-bench—sharply increased: scores rose by 18.8, 48.9, and 67.3 percentage points, respectively. Beyond benchmarks, AI systems made major strides in generating high-quality video, and in some settings, LLMs even outperformed humans in programming tasks with limited time budgets.

Note:
MMMU is a novel benchmark meticulously designed for college-level multi-discipline multimodal understanding and reasoning, and curated to assess the expert-level multimodal understanding capability of foundation models across a broad scope of tasks.
GPQA is a challenging dataset of 448 multiple-choice questions written by different domain experts, which are high-quality and extremely difficult. Experts who have or are pursuing PhDs in the corresponding domains reach 65% accuracy, while highly skilled non-expert validators only reach 34% accuracy, despite spending on average over 30 minutes with unrestricted access to the web (i.e., the questions are "Google-proof")
SWE-bench is a benchmark for evaluating LLMs on real world software issues collected from GitHub

A short 🧵about @swarmnode, as $SNAI keeps hitting ATH and is becoming another major AI agent tokens
👇👇 @swarmnode 1. What is @swarmnode?
It is a cloud infra for devs to deploy AI agents on-chain without any complex setups, due to several key features and benefits:

Here comes the most comprehensive 🧵of everything you need to know about @sporedotfun by @marvin_tong, whose $SPORE has been surging continuously to ATH during Christmas! In particular, @pumpdotfun and @sporedotfun have 2 things in common which makes me very bullish on $SPORE 1. TLDR: 2 things @sporedotfun & @pumpdotfun have in common are
- Wealth effect: with numerous chances of launching AI agent tokens (in different generations), @sporedotfun can keep a lot of people fascinated with their ‘get-rich-quick’ mentality, especially with the anti-sniper mechanism (explain it below)
- Speculative randomness: users in general are obsessed with the adrenaline rush caused by pitfalls - like lottery, but AI agent launch is even more speculative and random: each lottery is homogeneous, but each AI agent token is different in terms of traits and airdrop strategies (eg market cap (MC) divergence between $DREAM & $squiAId), with random mutations embedded in their 'DNA', simulating evolutionary forces in the natural world

Knowing that some particular AI agent tokens will be launched at a specific time, you can try to ape earlier and can some multiples of your principal as the return - eg you could do so by aping $DREAM 10 hours ago, now worth $7M+ MC from ~$500k MC

Last time I talked about @arcdotfun before $ARC went 3x, ppl asked me what's next. Here it is: $SWARMS @swarms_corp by @KyeGomezB - a multi-agent framework that enables enterprise-grade collaboration among multiple LLMs scattered around the world to solve complex tasks.
🧵👇👇 1. To know how important it is for @swarms_corp to facilitate collaboration among different LLMs, we need to know limitations of AI agents working in silo, which are —
1. context window limits: single AI agents can only handle a limited amount of info, incapable of dealing with complex tasks
2. hallucination: AI agents could produce seemingly right but factually wrong outputs
3. single-task execution: most AI agents are single-purpose, and cannot handle different nature of tasks without significant of model reconfiguration
4. lack of collaboration: AI agents operate in silo, lacking synergy
5. lack of accuracy: each AI agent has limitation in algo, and hence may produce inaccurate results
6. slow processing speed: AI agents may process large datasets and generate outputs very slowly, unable to give real-time responses
(elaboration of @swarms_corp solutions to each of them below)

Remember before $GOAT @truth_terminal got pumped, @pmarca sent $50k to help it 'escape'? Now @pmarca potentially "would fund a serious effort" to help an AI agent whose token is only traded at $30M market cap!

The story of @yesnoerror $YNE by @MattPRD: AI agent + DeSci
🧵👇👇 @truth_terminal @pmarca @yesnoerror @MattPRD 1. In October of 2024 a research paper claimed that black cooking utensils were extremely poisonous. People panicked. It turns out there was a just simple arithmetic error in the paper (60 multiplied by 7,000 to get 42,000)! More for the story: nationalpost.com/news/canada/bl…

1. TLDR of @unichain:
- It is a general-purpose L2, NOT app-specific L2
- Fast finality of 250ms per block, enabled by pre-confirmations to users
- *Priority ordering verified via TEE (collaborating with Flashbots team!)
- $UNI can be staked, or delegated to be staked
🧵👇 2. Before diving deep, the followings are worth attention:
- Trusted Execution Environment (TEE) gets applied by top protocol for the first time! (actively advocated by Flashbots team)
- $UNI finally has some utility 😂
- Fat App Thesis vs Fat Protocol Thesis - in the long run, apps may will capture more value than protocols (ie chains) - read @divine_economy's thread for more details: x.com/divine_economy…, also usv.com/writing/2016/0…
- Value accrual on Ethereum - more $ETH FUD is expected (e.g. x.com/0xBreadguy/sta…)

The $crvUSD design is the absolutely perfect for those who’d like to get the leveraged exposure to $BTC, via looping $btcUSD, especially during the bull market. Here comes @LambdaFinance and I’m very bullish on its clear goal of accelerate $BTC capitalization.

TLDR:
1. Compared to perps, $BTC long position is harder to get rekt by random wicks on CEXs due to continuous liquidation mechanism (vs one-off single level liquidation)
2. $BTC holders need not worry about redemption of their $BTC long by random ppl (vs $LUSD or $mkUSD where people complain that their $ETH or LSTs positions are redeemed even the collateral ratio is very high)
3. It’s far cheaper to get leveraged long $BTC by looping $btcUSD than by buying perps (CEXs’ funding rate is extremely high, which is why @ethena_labs can offer such a attractive APR to $sUSDe)

$crvUSD by @CurveFinance has shown its robustness with the ingenious LLAMMA design. The high level of $BTC dominance means eventually there must be ways to capitalize $BTC to make it more capital efficient. Imo being collateralized for the minting of $btcUSD should be the best way out among all (eg collateral for @ethena_labs or others).

Why?
First, $BTC whales generally are very risk-averse, so that any protocol without a long period of forward-testing can be convincing for them to deposit their $BTC position into it.
Second, the product should only contain $BTC or its equivalence as many whales have no trust in other assets, and thus mixing $BTC with other collaterals creates other unacceptable risks to their $BTC holdings.
@LambdaFinance’s $btcUSD seems to suit the taste of whales most, because (I) as an official fork of $crvUSD, forward testing for almost 2 years with stress tests during bear market is the best security guarantee it can offer, compared to other products, (ii) it only contains $BTC or it’s equivalence as collateral.

Edges of $btcUSD:
Leveraged long $BTC via CEXs’ perps requires you to pay an annual rate of at least >50% during bull market, compared to <20% by looping $btcUSD. The cost structure that @LambdaFinance can offer to $BTC long is unparalleled.
Also, it’s always a headache for $BTC perp buyers to get liquidated by random down wicks overnight; they lose money even when they are right in the market direction. This won’t be the case if they loop $btcUSD instead: LLAMMA’s continuous liquidation design allows liquidation to take place in a price range instead of a fixed price point, so that if the price range is wide enough, it helps protect ppl’s $BTC leveraged position from intra-day volatility.
Most stablecoin protocols have redemption mechanism to facilitate the peg via arbitrage opportunities, so that collateral holders could get their collateral redeemed even though their collateral ratio is at a healthy level. @LambdaFinance users won’t have the same problem, because $btcUSD peg is maintained by the dynamic interest rate, allowing the peg the stay in the dynamic equilibrium around $1: when it depegs downward, interest rate will increase to force borrowers to repay, creating buying pressure to push up $btcUSD; when it depegs upwards, apart from the printing of $btcUSD, interest rate will decrease to encourage borrowers to loop, creating selling pressure to push down $btcUSD. There’s no reliance of redemption to maintain price peg. As a result, $BTC position won’t get affected.

Every bull market comes with a stablecoin summer. Catalyzing $BTC capitalization is definitely a huge narrative coming. I’m bullish on @LambdaFinance. I suggest anyone sharing same view as mine can stay tuned to their twitter update:

https://x.com/lambdafinance/status/1780614491342061892

Have been reading through bullish tweets about @Blast_L2

Imo it’s okay to be a speculative play, but not my conviction bet. Here’s why: TLDR:
- it has existential risks by placing heavy reliance on @MakerDAO & @LidoFinance
- staking reward will be reduced in the future, and Lido could be further affected somehow
- the way of generating native yields on L2 could be easily followed by current L2 market leaders

This 🧵is about what @CelestiaOrg actually does, in view of a voting launched by @DefiIgnas and @VitalikButerin's recent article about @ethereum L2s

To begin with, we need to know what data availability is. Data availability refers to the ability of nodes in the network to access and verify the data associated with transactions and blocks.

Celestia's objectives:
– To make the blockchain networks more decentralized, by solving the problem of state bloat, i.e. as the state of blockchain grows with more and more transactions, the cost of running full nodes is higher due to higher hardware requirement for downloading more data, leaving only more well-equipped full nodes in place and thus the networks more centralized
– To improve TPS by allowing much larger block size for each block: to avoid centralization, there is often a limit imposed on the block size which compromises throughput/ scalability

TLDR: if I understand it correctly, @CelestiaOrg could provide a solution to the impossible trinity of blockchains: decentralization, scalability and security, by allowing sovereign roll-ups to be scalable yet decentralized through using it as the DA layer

Mechanism in details:
Consensus nodes produce blocks
—> Full nodes verify the block content by downloading all block data to produce fraud proof for light nodes; light nodes can’t verify but simply keep track of block headers (summary of block data) and work based on fraud proof
—> As the state of the blockchain (information necessary to execute all transactions) grows with the scale of the blockchain, it becomes more costly to run a full node, since more and more data has to be downloaded in order to produce fraud proofs for light nodes to work, or else light nodes will get tricked by block producers to follow a wrong chain by accepting an invalid block
—> Most chains limit the rate at which their state grows by imposing a block limit, or else the requirement of running full nodes will be more stringent over time
—> With DA modularization (e.g. @CelestiaOrg), light nodes allow full nodes to produce fraud proof WITHOUT downloading all block data each time, since full nodes can simply recapture the full block data now after combining all random chunks of data requested by each light node from block producers via Data Availability Sampling (DAS) which relies on a data protection technique called erasure coding
—> With DAS, the more the light nodes, the higher is the block size limit as full nodes are now assisted by more light nodes which altogether can request a larger amount of random chunks of data from block producers
—> Even with the same computation capacity of full nodes, the larger the block size, the more the transactions are processed by each block, the more efficient the blockchain is and thus the low gas fee is required in peak hours as the congestion can be largely alleviated by more and more light nodes joining the DA layer

Advantages:
– To align scalability with decentralization: with more light nodes, the block size limit is higher, i.e. more transactions can be processed per block with same computation power
– To avoid gas spike at peak times, as congestion is alleviated with better TPS by way of allowing a larger block size for each block

Question: how much scale can @CelestiaOrg offer?
- DAS key property: the more data is collectively sampled, the same probabilistic availability guarantees can be provided for larger sums of data, so that blocks can be made bigger (i.e. support a higher TPS) with more nodes participating in the sampling process
- Tradeoff: the block headers grow in proportion to the square root of the block size; light nodes looking to have almost the same security as a full node's bandwidth costs vary with (√n, n = block size)

2 factors of scalability:
i.) How much data can be collectively sampled: the block size can go much higher as DAS can be performed by a large audience with limited resources (even a smartphone can do it!). Thus one can fairly expect the number of sampling nodes to be correlated with user demand, defining blockspace supply as a function of demand. As it can offer low stable fees as the user demand grows, state bloat issue can be relieved
ii.) Target block header size of a light node grows in proportion to √n (n = block size): likely offset by improvements in network bandwidth over time as per Nielsen’s Law of Internet Bandwidth – if bandwidth capacity of light nodes grows by X, Celestia’s DA throughput can grow by X^2 @CelestiaOrg @DefiIgnas @VitalikButerin @ethereum Feel free to correct me if I have any misunderstanding!
@musalbas @jadler0 @nickwh8te @JoshCStein @ercwl @Cryptocito @Jskybowen @0xCygaar @dferrersan

Main issue in LSD: decentralization vs APR maximization
By making validator set permissionless (vs whitelisting like @LidoFinance), $ETH staked in the protocol may be idle and not fully utilized - eg $rETH APR always < $wstETH APR

@fraxfinance's $frxETH v2 solves it
🧵👇👇 @LidoFinance @fraxfinance 1. What @fraxfinance does in $frxETH v2 is to make staking reward to stakers as lending rate, whereas allowing anyone to be validators, who are the borrowers of $ETH staked in @fraxfinance, by putting 4 $ETH as collateral in case of slashing.

LSD = another form of lending market

A🧵to explain the set up of this new $crvUSD/ $fFRAX pool in the midst of @CurveFinance’s biggest potential crisis

TLDR: the objective is to decrease CRV/FRAX lending pair utilization rate (UR) & interests charged on him by Fraxlend

But why Fraxlend specifically?
👇👇 @CurveFinance 1. Because there’s a Time-Weighted Variable Interest Rate (TWVIR) which adjusts interest over time (max=10,000%) - the higher the UR, the shorter the half-life (12 hours at 100% UR) that doubles the interest.

Can you imagine one day you can get rewarded to do trades/ swap token assets?

@CurveFinance nailed it: traders are now charged with NEGATIVE FEES on LLAMMA!

Curve has reversed the law of mechanics and GRAB more TRADING VOL via the grand scheme of launching $crvUSD

🧵🧵👇👇👇 @CurveFinance 1. TLDR 1: $crvUSD turns demand of leveraging with a reduced risk of hard liquidation into rewards paid to traders who facilitate (de)liquidation when swapping on LLAMMA.

It renders gas optimisation and fee reduction as ways of DEXes to seize vol from aggregators less meaningful

[🧵@CurveFinance is printing money🧵]
$crvUSD is now charging on average >7.5% interest rate on borrowers, but still they are willing to do so, as a way of holding the leveraged long blue-chip - $BTC/ $ETH, without suffering from the risk of hard liquidation

More details👇👇

@CurveFinance 1. $crvUSD maintains its peg via its dynamic interest rate and Peg Keepers (similar to @fraxfinance's AMO) which maintain the peg by minting or burning $crvUSD in its AMM pools.

🧵about one of the most important updates of @CurveFinance (apart from $crvUSD): Tricrypto New Generation (TNG)

TLDR: Curve could FLIP @Uniswap's trading volume as TNG offers better execution price than Uni v3 by optimising gas, drawing huge vol from bots & aggregators

👇👇 @CurveFinance @Uniswap 1. This thread will be split into 2 parts -

i.) [2]-[14]: how TNG update can potentially facilitate @CurveFinance to outperform @Uniswap in volatile asset trading, by optimising gas fees

ii.) [15]-[22]: recap of Curve v2 mechanism and comparison with Uni v3's CLMM

This 🧵is about 3 key issues in DeFi lending. If they are solved, it can unlock super huge upside of DeFi by attracting TRILLIONS of USD from institutions!

They are:
1. interest rate isn't stochastic
2. utilization-based lending is inefficient
3. credit risk isn't composable

👇 1. Implications of 3 issues:
Fixed income, insurance, IR market... things related to risk pricing can't come to fruition without institutional money, as institutions don't know how to price those risks when this market differs much from the existing institutional money market.

A short recap about sell pressure of $ETH after Shanghai upgrade/ Capella:

TLDR:
- worst-case: sell pressure = ~3.4% $ETH daily trading volume, lasting for 5 days, assume all validators choose partial withdrawal (PW) and sell all $ETH reward at once (accumulated since 2020)
🧵👇 1. There are 2 kinds of $ETH withdrawal:
- Full withdrawal (FW): validators removing a full node of 32 $ETH from the staking pool
- PW: transferring the excess (>32) $ETH (i.e. staking rewards accumulated since the start of staking function in 2020) to the withdrawal address

Regarding $USDC crisis, it's time to revisit the most stable stablecoin - @LiquityProtocol' $LUSD with:
- 0 governance
- fully decentralized collateral + UI
- instant liquidation
- lower liq. level than @MakerDAO
- no loss to holders since launch

How can it achieve all these? @LiquityProtocol @MakerDAO 1. TLDR: 4 limbs for @LiquityProtocol to maintain the peg
- Redemption
- Stability Pool
- Redistribution
- Recovery Mode

After the 🧵, you will know how it can allow a volatile asset to peg with USD, with no further product update since launch but no loss caused to holders!!

As I said:

“Recession is forthcoming:
1. Yield curve inversion of 2yr (3mth)/10yr has always been followed by a recession, no exception
2. Historically there has never been a time where recession did not occur for CPI >5% to come down”

SVB is likely to be the start of recession I mentioned this last late Sept and it is still in play

https://twitter.com/DeFi_Cheetah/status/1574068975197356033

LPs of AMM can never be fully compensated by the trading fee (tx fee), which is a compensation only for LPs' cost of capital; LPs suffer from GAMMA RISKS as well.

This 🧵 will dissect the risk profile of being LPs of AMM, especially important to LPs of @uniswap v3!
👇👇👇 @Uniswap 1. TLDR: providing liquidity to AMM, you are just as an option trader hedging delta dynamically for his long put position, except that he profits from longing volatility, while you take gamma risks without complementary rewards.

I'll start from the basics and you'll understand

🧵@eigenlayer is solving the most important issue in crypto, that is, D-apps' innovation is capped by infra's rate of innovation

Impacts:
- ETH price goes up SIGNIFICANTLY
- Make innovations in infra PERMISSIONLESS, eg more well-designed oracles, bridges, sidechains, etc
👇👇👇 @eigenlayer 1. Key issue 1: rate of innovation at the core layer (eg consensus, sharding, etc) is way lower than that at the application layer, since whereas D-apps are permissionlessly deployable on blockchains, core layer upgrades are only permissioned (eg major upgrades on @Ethereum).

Recently I find a new project that facilitates us to find early hidden gems in crypto!

On-chain AI-powered search engine for the crypto community: @_kaitoai, backed by GPT 3, the tech underneath @OpenAI’s ChatGPT!!!!

👇🏻👇🏻👇🏻 1. Have you ever experienced a headache that information in crypto is so fragmented that they are scattered in different websites and you feel dizzy even with the help of high quality tweets that gather those websites? This is where @_kaitoai comes into play.