Will 威廉姆斯

Building with Impossible Fi for better crypto world｜ prev: Nervos . Now focus on stablecoin and AI in crypto .

[Will 威廉姆斯 @RBCHI]

https://t.co/bjKuPqDPdw

[Will 威廉姆斯 @RBCHI]

@sayinsaying knowing Jawz since OHM mania ,actually he is a real and brave man to share this .
Believe that it can be back someday

[jawz @sayinshallah]

I’ve wanted to tell this story for years. Never had the courage. Here it is.

I turned a presale allocation into $80 million on $OHM. Today I have $500k left.

In 2021, I got a presale allocation in OlympusDAO, then aped heavily myself on top of it. The allocation got me in the door. My own conviction made me go all in. Staked everything. Watched it compound daily. By the peak I was sitting on $80 million.

Then I started spending like the money printed itself.

Private jets to Dubai because commercial felt beneath me. $40k weekends in Monaco. A garage full of cars I drove twice. Watches I never wore. I tipped $5k at dinners just to feel something. Every purchase was a flex for an audience that didn’t care.

The casino was worse. High limit rooms in Vegas and Macau. I’d lose $2 million in a night and laugh it off because the portfolio would make it back by morning. Until it didn’t.

When $OHM unwound, I didn’t sell. I doubled down. Then I leveraged. 5x, then 10x, trying to trade my way back to the peak. Every liquidation felt like a personal insult, so I’d open a bigger position. I wasn’t trading anymore. I was gambling with a different interface.

$80 million became $20 million. $20 million became $4 million. I told myself $4 million was still life changing money. Then I levered that too.

$500k. That’s what’s left.

Here’s what I learned the expensive way:

Unrealized gains are not money. I never had $80 million. I had a number on a screen and the arrogance to believe it was permanent.

Getting in early is a gift. I treated it like a skill. The allocation didn’t make me a genius. It made me lucky. I confused the two for three years.

Lifestyle inflation is a leak you don’t notice until the ship is underwater. The jets and cars didn’t kill me. The identity did. I became someone who needed to spend to feel like a winner.

Leverage doesn’t get you back to even. It gets you to zero faster. Revenge trading is just grief with a chart open.

Nobody at the table in Monaco remembers my name.

I’ve carried this story alone for years. Too embarrassed to say it out loud. But $500k is more than most people will ever hold at once, and I’m done pretending the past didn’t happen. The next decade is about building slow and keeping what I make.

If you’re up big right now, screenshot this. You’ll need it.

[Will 威廉姆斯 @RBCHI]

上週的 Computex 的一些心得

1. 因為 AI agent 的興起，目前 CPU 的需求會越來越大

(這次老黃提到Nvidia 推出的 Vera CPU 也在直面這些邏輯運算的需求）
這已經和傳統 AI 推理的最大需求是 GPU 有所不同，過去對 LLM的推理是一次又一次的大運算，這是 GPU 的主場。

但 agent 問世後， Agent 的工作循環是:
推理 → 解析輸出 → 呼叫工具/API → 處理回傳 → 寫入資料庫 → 決定下一步 → 再推理。
中間那一大串「膠水邏輯」(解析輸出 → 呼叫工具/API → 處理回傳 → 寫入資料庫 → 決定下一步)全是有序列的而且充滿分支的密集工作，這一部分需要的就是 CPU。

Agent 越多、工具呼叫越頻繁,GPU 運算佔整體 workload 的比例反而下降,CPU 和記憶體的壓力同步上升，這就是為什麼這波 AI 架構轉變反而拉動了 CPU 需求。

2. 人形機器人（Humanoid Robotic）是 AI 的下一波浪潮

如果你錯過了 LLM 到 AI agent 的浪潮，那麼 Humanoid Robotic 將會是下個好機會，主要的原因是目前 AI 的觸手仍舊限於軟體的世界，需要有 Physical 的物體讓 AI 的影響力繼續延伸，因此 Physical AI 成為興起的主題，其中又以人形機器人是最多人討論。

但人形機器人目前離大規模使用仍有一段距離，不管軟體或者硬體，如足夠精細的關節、手，能夠負重足夠大的重量，或者是充電等等都還是挑戰，因此還處於能夠有進入空間的領域。

3. 對短暫兩年的經濟情況感到悲觀

業界普遍對於人形機器人能夠量產的預期都是 2028 到 2029 年，除此之外，未來兩年除了對於 CPU、記憶體以及 GPU 等運算存儲資源有很大的需求外，並沒有什麼特別的新技術能在未來兩年問世，並且讓大家激動，產能和需求可能也已經 price in 了，同時全世界也處於一種產業收益分配持續不均、通膨高漲、並且戰爭威脅環伺的情況，因此本人對於宏觀經濟是看衰的

此外，我甚至覺得市場已經過熱，在 Nvidia Inception 的展位中，我甚至看到純軟體的 AI 公司，目前產品還未量產，收入只從顧客的顧問費和個別軟體開發中獲取，但是估值卻超過 1b 的獨角獸。

4. 對於 Nvidia Inception Program 對進駐專案的孵化和協助感到驚訝

Nvidia Inception program 是個專門面向會用到 Nvidia 的新創公司的專案，原本我以為只是讓這些專案能夠獲取一些 Nvidia 運算等產品使用的協助而已，結果想不到在其中，有個進駐的新創公司 MetAI ，竟然透過 Nvidia 的牽線，而獲得台積電這個客戶，為他們提供 CAD to digital twins 的服務，算是讓我對於 Incpetion Program 進駐的新創又多了幾分信心。

5. 對於加密貨幣技術仍舊感到信心

區塊鏈技術和 AI 本身完全不衝突，唯一會競爭的只有投資人的注意力和資金，技術本身甚至相輔相成，像 Computex 這樣的大會裡，甚至沒有人討論到 AI 治理等等主題，而都是專注在技術的升級和未來創新，我覺得這是好現象，代表一個潛在的需求仍未被注意，而未來的世界凡但需要資產確權，例如賣資料給 AI Model traning 等，或者 AI 需要進行資產交易，這些都將會是區塊鏈的強項，與其現在才跟風去炒作已經非常熱的 AI 題材，專注在未來技術或者自己原本的強項或許才是康莊大道。

最後偷偷說，我在今年八月中文圈最大的開源人年會給自己挖了一個坑，然後聽說這份 Proposal 已經被大會接受了，可能要開始騷擾更多懂相關技術的大大惹

[Will 威廉姆斯 @RBCHI]

@btcbabycow Could you add me there ? Thanks

[Will 威廉姆斯 @RBCHI]

@0xHaley_ @Bankless @TrustlessState @Lighter_xyz the MEV side yeah it is a real concern esp on Lighter the MEC concern has been eliminated on tech side .
But tbh there is not too many evident or research saying that how huge impact of the MEV on HL .

[Will 威廉姆斯 @RBCHI]

Everyone's talking about $LIT after the @Bankless 's @TrustlessState shilled .

Here's my honest take not a call, just a breakdown.

What Lighter @Lighter_xyz actually gets right?

• ~100ms latency vs Hyperliquid's ~700ms
• Zero fees for retailer
• No validator committee for block production(less centralized )
• Buyback rate: ~6.4% of MCAP annualized vs HYPE's ~4.0%
• No major unlock until Dec 2026(the next unlock of HL is June 29 with 4.4 % mcap Hype)

On paper, a cleaner setup than people give it credit for.

But let's be precise about the latency thing
100ms vs 700ms matters enormously for HFT desks and market makers.

For the average retail trader? You will never feel this difference. I am not sure if it create the room to arbitrage or not for program.

Narratives built on tech specs that only matter to algos are fragile.

The buyback story is real, but contextualize it

$LIT: $2.2M in 30d buybacks → 0.52% of MCAP
$HYPE: $53M in 30d buybacks → 0.33% of MCAP

LIT wins on relative efficiency. HYPE wins on absolute firepower — by 24x.

Also: HYPE's buyback machine is already absorbing its monthly unlocks (~730k HYPE bought vs ~400k–1M unlocked). Net supply pressure: close to neutral.

LIT hasn't faced this test yet. Dec 2026 is when that changes.

Good read from @hell0men to learn much more about this

The unlock picture

$LIT: 50% still locked. 1-year cliff from Dec 2025 TGE → big unlocks start Dec 2026, then 36 months of linear vest.

$HYPE: Only 16.85% locked. Monthly unlocks already in motion, largely offset by buybacks.

LIT's "no unlock pressure" narrative is accurate for now. But the back half of the story is 3 years of gradual dilution starting in 6 months.

Per-event unlock: 5.56% of MCAP for LIT vs 4.46% for HYPE. LIT's relative hit is actually larger — amplified by thinner liquidity.

Can check it on @RootDataCrypto

Where Lighter genuinely falls short

The zk-SNARK architecture is legitimately impressive and I would say probably Lighter is the most technically complex perp DEX ever built.

But technical complexity is not a moat in DeFi. Execution is.

Hyperliquid shipped HIP-1 through HIP-4 from spot, perp, anything to Perp by anyone ,prediction markets with zero major bugs.
And never ignore that CoreWriter is opening composability for the next wave of builders. The ecosystem flywheel is already spinning.

Lighter has the better whitepaper. Hyperliquid has the better traction and possibility on innovation and ecosystem .

The wallet integration gap

@MetaMask . @phantom . @Rabby_io for such mainstream wallet have integrated the perp feature powered by Hyperliquid . Even there are some wallet specifically for Hyperliquid such as @BasedOneX

User habits are sticky and the switching costs are real. Lighter needs more than superior tech to convince user to move.
That reason doesn't exist yet from my observation .

One more thing about the KOL pump

When Bankless starts covering a $400M MCAP asset, ask yourself: who is still left to discover this?
Narratives that go mainstream don't disappear but the easy money usually already moved.

My read

Lighter is a technically serious project with a real (if narrow) edge on tokenomics through 2026.
But "better in a spreadsheet" rarely beats "already has the users."
HYPE would need a serious self-inflicted wound such as exploit, compliance action, execution failure then there would be the oppurturnities for LIT to realistically close the gap.

I'm watching. Not concluding.

DYOR.

HyperLiquid

Some good material to learn more about Lighter
- David's post : https://t.co/bHTAjUBH2r

- Dashboard of Lighter from @Delphi_Digital @trevor_flipper
https://t.co/jeG77T5NtB

- Price discover performance comparison from @ArrakisFinance
https://t.co/skFB4OAUE6

[Arrakis @ArrakisFinance]

https://t.co/E8muV08EZz

[Will 威廉姆斯 @RBCHI]

@Levi_Researcher @Bankless @TrustlessState @Lighter_xyz yes not quite so many users will read and decide to use it after the reading
but actually I like to read the first hand doc somehow

[Will 威廉姆斯 @RBCHI]

@btcbabycow but what is the utility of the $aeon ?
I still didn't get it

[Will 威廉姆斯 @RBCHI]

Surprised that the lag performance of @Lighter_xyz is even better than @HyperliquidX
But not sure how exactly the user feel and how's the impact for high frequency traders ?

[Arrakis @ArrakisFinance]

https://t.co/E8muV08EZz

[Will 威廉姆斯 @RBCHI]

在這個 AI 能做東西的年代裡
用 LLM 搞出花拳繡腿的作品已經很難吸引注意
所以產品能不能直擊用戶的心臟就很重要

[Will 威廉姆斯 @RBCHI]

Echoed by @eigencloud as what I rmb that EigenCompute also take lots of effort on deterministic computation for LLM but not sure what if the huge AI giant such as Nvidia also focusing on it . How can those web3 company respond .
@sreeramkannan Sreeram I would like to learn much more from you about this

[Will 威廉姆斯 @RBCHI]

What is the "Hardness" that Jensen from Nvidia said in the GTC presentation ?

One concept is forming from today for AI computation : Hardness .
Hardness isn't hardware performance but it means "determinism" whether an LLM can produce reliable, verifiable, zero-deviation output when facing structured tasks.

Why the hardness is the bottleneck: Models keep getting smarter but less reliable. Ask the same question twice and get two different answers ,we can say this feature in conversation is a strength cause much more diversity,but it is a fatal flaw for Agents if it is not reproducable .

No bank, hospital, or law firm will hand critical tasks to a system that's only "probably right" cause there is no room and zero tolerance for 1 mistake happen on AI .

So there are two solutions so far
1. Constrained Decoding : adds a "grammar firewall" during generation, mathematically guaranteeing 100% format correctness
2. System layer validation : Agent self-checks before executing, cross verifies via SQL/API calls

And here one thing Jensen hide is that agentic AI is making SQL and relational databases critically important because Agents need "Ground Truth." SQL's ACID properties provide the determinism is exactly what LLMs lack most .Yeah Echoed by Vera CPU's 3x SQL performance.)

And here Jensen also show what the Nvidia's hardware play to build the infrastructure for the "Hard Agent" Era

1. Vera Rubin (now in full production): not a GPU, but a complete agentic computing system. Grace Blackwell was built for inference, Vera Rubin is built for agents

2. Vera CPU: first CPU designed for Agents, living in the "nanosecond" world with SQL 3x, streaming 6x

In the future ,there is not just GPU from Nvidia but all the computation related business from GPU to specific CPU to Agent.

[Will 威廉姆斯 @RBCHI]

@fluixoo absolutely it is

[Will 威廉姆斯 @RBCHI]

@jocyiosg
突然想起來，第一次見到 Jocy 是 2019 年在杭州，那時候我還在 Nervos ，轉眼間現在亞洲的 VC 很多能叫出名字的都沒了，但 IOSG 確實還在研究，建設以及投資，這就是會被週期獎勵的人，也是真正值得 respect 的從業者

[Joe Zhou @joezhoublack]

https://t.co/dqJ5hYcXze

[Will 威廉姆斯 @RBCHI]

@Okada_DeFi0x @HarukoRose If the usdm is from team’s fund then it is a good move for me at least less damage for community interest

[Will 威廉姆斯 @RBCHI]

中文直譯：
看了一下我們的 Terminal data 發現用戶的真實參與情況並不咱地，所以我們把 Terminal 關了，並且用 USDM 負責獎勵參與人。

------

說真的，看 Terminal 的頁面，怎麼看都不像是只有要 run 兩三周的活動（辛苦 @HarukoRose )，但現實情況如此確實要馬上應對，不然代幣就打水漂了

結論：少了 sell pressure，利好 Mega

[MegaETH @megaeth]

After reviewing the Terminal data, we’ve decided to conclude the season early, and sunset our Terminal program.

The points distribution for Week 3 has gone out as normal and a snapshot has been taken of all user activity up to this point, which will be used in our final calculations. This decision did not come easy, but we believe this is what’s best to support our ecosystem.

Next week, we'll run a supplemental points distribution to thank all users who have been exploring and experimenting with the different apps throughout the program.

Eligible Terminal participants will receive a share of the rewards pool in USDM, based on the points accumulated to date as well as activity that supported Mega.

Early next week, we'll open a place on Terminal for one week so you can designate a wallet address to receive your rewards.

Terminal’s main features will be merged with Rabbithole, transitioning everything into one unified chain experience.

Thank you to everyone who has supported us this far. We can’t wait to show you what we have in store.

A note from Shuyao on what’s next:

[Will 威廉姆斯 @RBCHI]

Actually it would be a bit arrogant when a decentralized layer2 project doesn’t take the communication with community on the biggest platform you exposed as an important thing . Unless the program you are working overtime everyday can hit a home run ,or it must lead some concern.

[Will 威廉姆斯 @RBCHI]

@cloister_space Two suggestions
1. Open the community chat
2. Open source the code on github or @gitlawb

[Yueya (ARX Mode) @yueya_eth]

很认可 @thedefiedge 对 Base 上 robotics × crypto 的梳理。

我自己目前最看好的，还是 $CAS (@caspius_ai )。

原因很简单：

现在大家都在聊 AI agent、机器人、physical AI，但真正卡住行业的，其实不是模型，而是「embodied AI data」。

机器人不是会聊天就够了，它得知道：
怎么抓东西、
怎么移动、
怎么在真实环境里完成任务。
而这些东西，本质上都需要海量第一视角的人类行为数据。

我觉得 $CAS 很特别是因为：它把“普通人的日常动作”变成结构化训练数据。

做饭、洗衣、整理房间这些过去毫无价值的行为，现在都能被记录、标注、训练、货币化。

Genesis NFT 的设计我也觉得很聪明：

• NFT = 数据运营商席位
• 不同身体部位 + Traits 决定专精方向
• 稀有度 / Level / Trait match 会影响 multiplier
• Fabrication 可以持续合成、繁殖、进化“血统”

本质上，它是在把数据网络游戏化。

你不只是持有 NFT，
而是在参与训练未来机器人的“动作数据库”。

最关键的一点是：
很多 robotics 项目做的是 execution layer、coordination layer、payment rails。

但 $CAS 抓的是更底层的东西：数据。

而数据一旦形成飞轮，护城河通常会比协议本身更深。

但至少目前看，它是整个 robotics × crypto 里，少数让我觉得“方向对了”的项目之一。

[Will 威廉姆斯 @RBCHI]

@btcbabycow Ui 很像是Claude 做的

[Will 威廉姆斯 @RBCHI]

@yueya_eth 我想和你一起淋雨❤️