Will Sutton

🛡️Gauntlet - Cohort 3
- Innocuous: Agro: LightningBounties @LBounties

[Will Sutton @WillSuttonCodes]

@pensandpoison Frost's 'Mending Wall' may be instructive here:

"Good fences make good neighbors" is repeatedly spoken by the antagonist.

Contrast this with protagonists language:

Bringing a stone grasped firmly by the top
In each hand, like an old-stone savage armed.

[Will Sutton @WillSuttonCodes]

Relevant advice in the age of LLMs. Cliches are the 'most likely next token' in fiction.

Good writing (and good humor!) Is about defying expectations, while still maintaining the internal logic of the text.

This may explain why AI still cant create good jokes or compelling fiction.

[How I Write Clips @PerellClips]

"You get at most one cliché per book."

Jonathan Franzen doesn't finish most books people send him. He reads until the second cliché, then stops.

"People send me a galley of a book that's coming out. I read along until I get to the second cliché, and I say, thank you."

"They tell me this is not an attentive writer. Aren't you aware that that's a cliché? Didn't it occur to you to stop and take 20 minutes to find a better phrasing than that?"

"You can also have borrowed feelings, borrowed ideas, borrowed situations. Sentiment can be clichéd, or unearned."

[Will Sutton @WillSuttonCodes]

@planefag This time might be different...
Did you hear Cougar just turned in his wings?

[Will Sutton @WillSuttonCodes]

@DrBitcoinMD Hero

[Will Sutton @WillSuttonCodes]

@alexcdot I implemented the watermark algo in python / llama.cpp

https://t.co/WJriIfBSdv

[Will Sutton @WillSuttonCodes]

I implemented the watermark algo in python / llama.cpp

https://t.co/WJriIfBSdv

[Jonas Geiping @jonasgeiping]

Why am I being baited by watermark misinformation on this app, is it 2023 again?

A small FAQ:

1. 𝗪𝗵𝗮𝘁'𝘀 𝗮 𝘁𝗲𝘅𝘁 𝘄𝗮𝘁𝗲𝗿𝗺𝗮𝗿𝗸? -- A modification of the LLM sampling algorithm that, if there are multiple ways to write something, will pick one that agrees with a pseudorandom key. This is a local, invisible signature hidden in the way phrases are used in any LLM text that persists when text is copied.

2. 𝗗𝗼𝗲𝘀 𝘁𝗵𝗶𝘀 𝗺𝗮𝗸𝗲 𝘁𝗵𝗲 𝘁𝗲𝘅𝘁 𝘄𝗼𝗿𝘀𝗲? -- A good implementation is 'undetectable' (in polynomial time), meaning: If you do not have the private key, then neither you, the model itself, or pangram could detect that this is happening.

3. 𝗪𝗶𝗹𝗹 𝘁𝗵𝗶𝘀 𝗯𝗿𝗲𝗮𝗸 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹'𝘀 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴? -- Because Ant already encrypts the model's reasoning, they can just not watermark the model's internal reasoning, leaving the thinking unaffected.

4. 𝗪𝗶𝗹𝗹 𝘁𝗵𝗶𝘀 𝗺𝗮𝗸𝗲 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝗹𝗲𝘀𝘀 𝗰𝗿𝗲𝗮𝘁𝗶𝘃𝗲/ 𝗺𝗼𝗿𝗲 𝘀𝗮𝗺𝗲-𝘆? -- If anything this (marginally) increases entropy across different generations, so it will make model outputs slightly more varied.

5. 𝗕𝘂𝘁 𝗜 𝗰𝗮𝗻 𝗷𝘂𝘀𝘁 𝗿𝗲𝗺𝗼𝘃𝗲 𝗶𝘁 𝗽𝗮𝗿𝗮𝗽𝗵𝗿𝗮𝘀𝗶𝗻𝗴? -- Absolutely! But, judging from the amount of writing on the web that already unmistakably sounds like Claude, most people likely will not bother.

5b: Also, not any paraphrase will work. To remove (for example) a k=5-minhash watermark completely from a long document, you need to make sure none of the original 2-grams, 3-grams, 4-grams, 5-grams and 6-grams of the text remain.

6. 𝗪𝗶𝗹𝗹 𝘆𝗼𝘂 𝗶𝗻𝗮𝗱𝘃𝗲𝗿𝘁𝗲𝗻𝘁𝗹𝘆 𝗰𝗼𝗽𝘆 𝘁𝗵𝗲 𝘄𝗮𝘁𝗲𝗿𝗺𝗮𝗿𝗸? -- No, with a good implementation the space of possible realizations of the key is too large to memorize.

7. 𝗪𝗶𝗹𝗹 𝘁𝗵𝗶𝘀 𝗮𝗹𝗹𝗼𝘄 𝗖𝗹𝗮𝘂𝗱𝗲𝘀 𝘁𝗼 𝗶𝗱𝗲𝗻𝘁𝗶𝗳𝘆 𝗼𝘁𝗵𝗲𝗿 𝗶𝗻𝘀𝘁𝗮𝗻𝗰𝗲𝘀 𝗶𝗻 𝗮 𝘀𝘄𝗮𝗿𝗺? -- The watermark will 'appear' like random sampler fluctuation to the model and would not be detectable. But, if an agent gets hold of a detector endpoint, it can absolutely use the watermark to ID other Claude agents (not that it would have trouble noticing them based on their writing as of today).

8. 𝗪𝗶𝗹𝗹 𝘁𝗵𝗶𝘀 𝗱𝗲𝘁𝗲𝗰𝘁 𝗱𝗶𝘀𝘁𝗶𝗹𝗹𝗮𝘁𝗶𝗼𝗻? -- By default, no. If the watermark is set up to be 'undetectable' (as assumed above), it will not be picked up in training by other models. For that to happen, the watermark needs to be detectable by ML algorithms.

9. 𝗪𝗶𝗹𝗹 𝘁𝗵𝗶𝘀 𝗺𝗮𝗸𝗲 𝗣𝗮𝗻𝗴𝗿𝗮𝗺'𝘀 𝗷𝗼𝗯 𝗲𝗮𝘀𝗶𝗲𝗿? -- By default no, this is a separate avenue to detection. But, they might collaborate with Anthropic which would allow them to detect the watermark as well and show a watermark score next to their text detection score.

10. Bonus: All aside, is this a good idea? I don't know. The companies are doing it to follow the writing of the EU AI act, which was written based on 2024 information and when the field looked very different, and threat models were focused much more on slop/propaganda (like the Kokotajlo 2026 prediction). The actual 2026 looks quite a bit different.

[Will Sutton @WillSuttonCodes]

Turns out Factorio was correct.

This is how you "science" in the future and expand your tech tree.

Deliver a huge belt of resources into glowing buildings and let them invent.

[Dwarkesh Patel @dwarkesh_sp]

"Lab compute 3x-es year over year. For a lab to 10x revenue while continuing to only 3x compute, some combination of the following 3 things has to happen:

1. Lab margins have to increase,

2. The price of compute has to increase,

3. Labs have to spend a greater fraction of their compute on inference.

My understanding is that basically all 3 of these things have been happening..."

"Consider the price at which Google and Anthropic are renting compute from SpaceX.

Google is reportedly paying $900 million a month for 110K GPUs [at] ... roughly 2x the spot price per hour for those GPUs. And the current spot price is itself 40% higher than it was in February."

[Mike Baier @Mike__Baier]

I’ve had quite a few comments asking me to provide actual pictures of the loop instead of AI slop. For the record - this is a real picture 😂

[Will Sutton @WillSuttonCodes]

@jmrphy Thesis: market for text documents trifuricates
a) Prestige media, unique authorial style e.g. "Infinite Jest"
b) Comfort slop, possibly re-written in the consumer's preferred style, e.g. Romance books, anime
c) Clerical work, standardized style, e.g. corporate employee handbook.

[Will Sutton @WillSuttonCodes]

@bayeslord Its no longer about IQ. Agency is all you need.

And war is its greatest test harness.

[Will Sutton @WillSuttonCodes]

@pesottas Functionality does not rise linearly with accumulated unrepaired damage?

Instead it may act as sharp sigmoid where a straw breaks the camel's back. Moving the organ system from functional to non functional in a short amount of time.

Would this not explain the phenomenon?

[Will Sutton @WillSuttonCodes]

This scheme actually works for @kaggle competitions

https://t.co/RKaUsRTprK

[Will Sutton @WillSuttonCodes]

Undertaker son or over-sharer daughter?

[Will Sutton @WillSuttonCodes]

I thought the Bears were moving to Indiana 🤷‍♀️

[New York Post @nypost]

Mass grizzly bear reintroduction in California moves step closer https://t.co/SkntDv3pwR

[Will Sutton @WillSuttonCodes]

@zriboua There must have been 1000% increase in short beards for 20-40 year old middle class men.

Happend between 2000 and 2020. Theres an interesting cultural explanation too.

People will notice that the mustache is now taking over since 2024.

[Fred Krueger @dotkrueger]

My take on the Clarity Act:

1. For Bitcoin. Very Bullish. Self Custody explicitly protected. Clear legal framework for lending, wrappers etc.. Banks can go nuts.

2. For DeFi. Generally Bullish. Protocols are intact as long as they are decentralized. Front ends need to do more Geo Blocking / SAR / potentially KYC.

3. For Stablecoins. Bullish, but yield bearing coins get heavily restricted. Banks win.

4. For "Crypto/Bitcoin Companies". Very Bullish. US Companies building truly decentralized protocols are fine. Products can start out more centralized and decentralize to comply.

This would start being enforcable in summer 2027 according to Claude.

[Will Sutton @WillSuttonCodes]

Cool demo from @OpenAgents

Sell your compute for bitcoin!

https://t.co/FVc7S1I5RS

[Pirate Wires @PirateWires]

We interviewed @charliebcurran, the AI filmmaker behind these legendary Spencer Pratt videos — while Hollywood panics about the technology, Charlie’s optimistic: “[AI] is how you get an artistic revolution,” he told us.

Full profile from @eventidia below 👇

[Will Sutton @WillSuttonCodes]

Youre right to challenge the narrative "Amateur solo dev bests entire industry".

It was a prediction for ai impact and didnt show up yet.

In 2020s america, talent is allocated to software like brazilians cultivate soccer talent.

Thats tough competition for an unfunded amateur.

[Will Sutton @WillSuttonCodes]

@xwanyex Dont overindex on
all tech = consumer entertainment

Plenty of deep tech work is behind the scenes, e.g. the "shale revolution" (aka fracking) is the only reason our home energy bills havent sky rocketed and why we can build and power the data centers.