building @pumplyco⛽️ looking for the bugs in people, systems, and ideas.
at the pump
Joined December 2025
- Tweets1.9K
- Following728
- Followers550
- Likes17.8K
Clif retweeted
Today I call upon the blockchain industry to calmly begin planning for "bunker mode". My personal recommendation is to set in motion a controlled mass migration of assets to fresh addresses, i.e. addresses whose pubkeys remain hidden behind a hash.
Holders, starting with large and sophisticated ones, should consider moving the bulk of their funds to addresses that have never signed a transaction. And when they do sign one, they should also move remaining funds to a new address (possibly generated from the same seed phrase).
Don't rush. While I believe there is cause for action a rushed migration would do more harm than good. Don't panic either. Moving assets to protected addresses is a simple, preventative step which does not require new cryptography or new wallets.
IMO it is now reasonable to brace for the possibility that ECDSA breaks before qday, in the worst case in months not years. By "break" I mean fast private key recovery (e.g. in one week) on available hardware (e.g. a large GPU cluster).
Recent days have been humbling for human mathematical intuition. Long-held, unquestioned hypotheses have fallen. This includes the n log(n) bound for integer multiplication and the 3SUM conjecture. In hindsight, May's unexpected disproof of the Erdős unit distance conjecture was our warning shot.
Yesterday's OpenAI drop made it clear that mathematical superintelligence is upon us. They say there are weeks where decades happen. We are about to live through weeks where centuries of mathematical progress happen. Could our magic 64-byte ECDSA signatures be too good to be true? Was it just security through obscurity all this time?
Elliptic curves feel especially vulnerable to superintelligence. Curves carry rich structure, with room for fancy tricks like Schoof, Frobenius, pairings. (By contrast, hashes are designed to minimise algebraic structure.)
Separately, as Ewin Tang can attest, an efficient quantum algorithm sometimes foreshadows an efficient classical one. We should be open to the possibility of a classical counterpart to Shor that breaks elliptic curves and RSA at once.
Also noteworthy is the striking under-representation of cryptographic breakthroughs among the 722 mathematical results OpenAI published. I've witnessed first-hand the US government censoring academic quantum cryptanalysis results. Backroom interventionism is my base case.
I urge large, sophisticated actors to lead by example. Project11's "risq list" (bitcoin-risq-list.projecteleven[.]com) is a great tracker of exposed BTC pubkeys. Binance, Bitbank, Robinhood, Bitfinex, and Tether have an opportunity to harden their cold storage. Next month I'll address institutions in London in a live Q&A (forum.ethereuminstitutional[.]org/london-2026).
Again, please do not rush. Wallets holding under 50 BTC enjoy partial cover from "Satoshi's shield", i.e. his 20K exposed addresses that hold 50 BTC each. Load-bearing signers like oracles and L2 security councils should consider rotating ECDSA pubkeys with every signed message and/or multi-signing with a hash-based schemes like SPHINCS.
Exiting bunker mode safely will require post-AI cryptography. My inclination is to go all-in on hash-based cryptography and avoid structured mathematical assumptions entirely, whether from curves, lattices, or isogenies. A single battle-tested hash (e.g. from the SHA or BLAKE families) yields plausible post-AI security.
The Ethereum roadmap on strawmap[.]org fully embraces hash-based cryptography with end-to-end formal verification as a response to the quantum threat. Those timelines must now be revisited and accelerated in light of mathematical superintelligence. I'll be pushing for maximum defensive acceleration.
Clif retweeted
Chat you soon! 👇
part 2 with @Dawoodiddris tonight. part 1 ended right when it was getting good, so we're picking it up from there. 8:30pm, pull up 🎙️
nitter.cf/i/spaces/1qJDzWNaozBKV
Clif retweeted
JUST IN: 🇨🇳🇺🇸 Chinese businesswoman with close ties to Nvidia CEO Jensen Huang identified as spy secretly passing information to China's government.
Clif retweeted
AI is allowing me to do more.
Doing more is making me tired more frequently.
Before AI, I was already a bit of a workaholic. So imagine what the situation is now.
I'm constantly having to remind myself to stop and just be. Rather than to keep pushing past my fatigue limits.
Clif retweeted
@seevcash is headline sponsoring Google Developer Festival Accra (@GDGAccra)
Over the years, Google DevFest has been one of the places where Ghana’s builders show up, learn, and leave sharper.
Proud to back that room this year. Spot our booth, meet the team, and if you’re coming, bring a friend onto SeevCash before we gather under one roof.
Before we gather under one roof 👀 make sure your friends are not left out.
Bring a friend onto Seevcash and you could both get $10 on your Visa cards 💳
Here is how it works:
They sign up on Seevcash
They transact over $500
You get $10 on your Visa card, and they get $10 on theirs too 🎉
More friends, more rewards. Share your referral code and bring your people with you.
#Seevcash #ReferAFriend #VisaCard #BorderlessPayments
Clif retweeted
AI is heaven for coders because it automates what they don't want to do, and hell for mathematicians because it automates what they do want to do. Every profession will fall somewhere on that spectrum.
Clif retweeted
i honestly did not think openai would expose dots to imessage for a variety of reasons but since the competition is so so steep now particularly from meta, it makes a ton of sense. meta is obviously working very hard to bring all of muse to imessage since whatsapp integration already exists.
that's actually hugely validating for my earlier post about imessage. i think there is a really good chance apple has to embrace the imessage agent phenomenon at scale by supporting it.
also it's fairly clear now that openai's hardware device is a dot that moves around your home, like a portable homepod but with eyes & a screen for generated ui like what's in the chatgpt app now. then there will be a portable dot with "airpods" with cellular connectivity, one that lives in your pocket. the hardware roadmap could not be any more clear.
Clif retweeted
you know if you really think about it apple may have accidentally stumbled upon a huge new business with imessage agents.
imessage is a potentially huge ai platform opportunity for apple in some sense (assuming this behavior actually works in the long term which is still a big non trivial if).
imessage apps mostly didn't hit, but imagine a store of agents that apple curates & allows? imagine the api's apple could facilitate here. accessing health through imessage agent securely & safely? no problem. accessing contacts? no problem. accessing location in an agent friendly way? no problem. payments for agents? no problem. subscription payment for using agents? no problem.
imessage agents could be the new app store if you will for discovery & access for both developers & users, just like the original app store. if jobs were alive, you can kind of picture the keynote for this. lots of implications/nuances for both iphone & apple's overall business but obviously very fun to think about.
this should be a topic of discussion for apple leadership if not already. there is a giant moat here for apple that it can uniquely make work.
Clif retweeted
For anyone using Grok Bot or Claude to make video edits, there are a few things nobody seems to be doing.
This is a small clip from an upcoming video of mine.
FIXES: graphics pinned to the screen instead of the scene, dark scenes that break into bands and blocks on X and YouTube, numbers that jiggle while they count, and soft shots that don't match the rest of the edit.
👾Graphics pinned to the screen instead of the scene
The top half of this video has its graphics pinned to the screen the way most edits do, and the bottom half has them locked to the world.
When the camera moves, a label that points at something stays where it was and the picture slides out from under it. NASA Ames tested this with pilots in 1995. Landing in a simulator, they took 2.5 seconds longer to react to a plane on the runway when their flight symbols were on the head-up display instead of the instrument panel. Their explanation was that symbols on the glass don't move the way the world outside does, and that difference pulls attention onto them. Their fix was to draw the symbols as objects in the world so they move with it, and pilots flew their path more accurately.
In an edit, that means tracking how the background moves in each shot, ignoring the people in it, and moving every graphic that points at something along with the camera. A label on a target follows the target, and titles and counters stay fixed.
👾Dark scenes that break into bands and blocks on X and YouTube:
X and YouTube compress what you upload, and the smooth gradients in dark footage turn into bands and blocks. The fix is to smooth those gradients before the grain goes on.
👾Numbers that jiggle while they count:
Counters jiggle because the digits in a lot of fonts are different widths, so the number shifts sideways as it counts. Putting each digit in a fixed-width slot keeps it still.
👾Soft shots that don't match the rest of the edit:
Measure the sharpness of every shot too. The softest one stands out next to the shots around it, and that's the one to regenerate.
Clif retweeted
internet native communities in ghana already coordinate in real life
2021: a hashtag put people on the streets of accra. now: spaces and group chats raise money, find jobs, organize meetups
network states will start wherever the group chat is stronger than the institution
Clif retweeted
best breakdown of ghana's Ai plan i've read, and the numbers are the story.
one data centre = 90% of the country's entire commercial data-centre load, costing 6x the industry's revenue to run. that's not building on a market, that's outsizing it
1. Readers of this page know by now that JoyNews held a forum on Ghana's AI strategy and prospects. And the Digital Minister and I got into a serious debate about his Ministry's plan to build Ghana's competitor to Gemini, Kimi, ChatGPT, Qwen, Claude, etc.
2. Ghana's competitor to these global AI platforms is called GhanaChat.
3. It will be trained and hosted in a brand new AI datacenter.
4. 48 registries pooling data from across the public sector will make this AI smarter about Ghana than the foreign ones.
5. The AI datacenter and the "AI Fund" will cost about $270 million. Expect more to be spent on the registries and the apps. A lot of contracts. You can bet that a lot of ICT contractors are very happy.
6. The question is, if we have a couple of hundred million dollars to spend on making AI deliver great benefits for Ghana, should we spend it on GhanaChat, a top-tier AI data center, and the data-pooling system? And should we do it according to the strategy the Ministry has devised?
7. Or could we do much better by more critically assessing all our strategic options and comparing them to the Ministry's?
8. My argument on the show was that I am not convinced about GhanaChat offering the best value for our money.
9. It would take many pages to explain my position fully. But in this post I will highlight just the most important points.
10. Attached are extracts from page 13, 50, and 68 of Ghana's AI strategy.
11. If you have even the barest exposure to AI, you can see a massive error. Can you?
12. The target for how much computing capacity the country needs for next year (2027) is indicated as "1021 FLOPS" and the 2035 target as "1025 FLOPS".
13. If this was indeed correct, it would mean that the Ministry would like the entire computing capacity of the country in 2027 to be equivalent to a 1970s desktop computer. Or 2% of the capacity of the 1980 Intel math chip.
14. Howzat?
15. I'm sure you can agree that this would be totally ridiculous if that was the Ministry's intent. So, what happened?
16. Somebody who didn't fully understand how FLOPS work must have been involved in the drafting. What was meant to be 10 raised to the power 21 operations per second (10²¹ operations a second) was written out as "1021". Same for "1025".
17. Why wasn't the error caught? Because in Ghana when doing "consultations", governments rarely engage independent and critical voices who will point out absurdities without fear of offending big men and women.
18. I am not being petty. There is a serious strategic confusion beyond mere typos.
19. Read properly as 10²¹ FLOPS, the Ministry is telling us that by next year it will have some 200,000 of Nvidia's newest chips in use in Ghana, 500 MW of power devoted to AI, all costing about US$24 billion (close to the Ghanaian government's entire programmed spending for 2026).
20. The US$250 million AI datacenter now being procured supplies about 1% of this 2027 target.
21. Howzat?
22. Clearly, no serious critique of the plan would have okayed such a vision. Throughout the ministry's own review, cabinet approval, and a high-profile media launch, it is absolutely clear that not one critical reviewer was engaged.
23. My first major point here is that this is classic KATANOMICS: the lack of critical policy audiences. This is exactly how a sugar factory that was "designed from the start to fail" was allowed to be built. Many programs and initiatives in Ghana fail for precisely this absence of critical review.
24. In the strategy, you find statements like: "government aggressively building AI compute clusters" and by courting hyperscalers, "to leverage the scaling laws towards building general purpose AI and AGI". Then: "drastically increase Ghana's energy capacity 0.5 to 1 OOM per year". "OOM" means "order of magnitude".
25. When I challenged the Minister on TV, he confidently rebutted by claiming that the Energy Minister/Ministry has assured him of feasibility. Ghana's energy Ministry is committing to increase Ghana's power output by 3x to 10x EVERY YEAR so that the Digital Minister can deliver his AI strategy? Did anyone in Cabinet not protest that this projection in 8 years would mean expanding Ghana's electricity output by 10,000x?
26. I can bet you any amount that the Energy Minister will deny making any such commitments in Cabinet. Can we have clarity? Can JoyNews check?
27. Page 68 of the strategy defines an "Artificial General Intelligence Lab" as "an AI lab with access to more than 1010 compute". No unit is given at all. In fact, as written it is totally meaningless. What they likely mean is that Ghana will support the establishment of a lab with computing capacity of 10¹⁰ FLOPS (10 gigaFLOPS).
28. The problem is that such a lab would be totally useless. To give you perspective, an iPhone 7 has about 350 gigaFLOPS of capacity.
29. What would be required to meet the Ministry's 2027 AI target? If for frontier model development, then the following spec would be reasonable: 400,000 16-bit Nvidia B300 GPU units, 1,114 MW of dedicated power (which is 26% of Ghana's peak demand). And the budget could easily hit US$47.2bn. For less frontier use, we can halve these numbers.
30. Counted in the 64-bit arithmetic that scientific supercomputers are ranked by, Ghana's 2027 target is 455 times the world's fastest machine, China's LineShine at 2.2 exaflops, and 53 times all 500 machines on the list combined.
31. Anyone with the barest exposure to Ghana's fiscal context can tell the Minister for free that this won't happen next year.
32. Ghana's 2035 target of 10²⁵ is 160 times the entire global stock of AI compute today. Even if chips grow 1.4 times more energy-efficient every year, it would draw about 377 GW, 65 times Ghana's current installed generating capacity.
33. I will be charitable and propose that the strategy's drafters may have been thinking about training FLOP (note: without the "per second" at the end). For example, GPT-4's training run used about 2.1×10²⁵ FLOP, the same order of magnitude as Ghana's 2035 endpoint.
34. Unfortunately, that doesn't save the strategy. Recast this way, the target collapses the other way. 10²¹ FLOP is about 12 days' work for one H100 chip at 16-bit precision. 10²⁵ FLOP is one GPT-4-sized training run, which the US$250 million centre could finish in about four months. On that reading the centre meets the 2035 target before its first anniversary. Not meaningful.
35. Given how confusing the overall strategic plan is, would you bank on the first AI compute center and GhanaChat product already being procured at about $270 million?
36. Here are some bare numbers: the budget, per our modelling, maps to about 2,100 GPUs and 4.3 MW of IT load. That alone adds the equivalent of 90% of Ghana's entire commercial data-centre load in 2024 (4.8 MW). Running it costs about US$76 million a year, which is about six times the revenue of the whole commercial industry.
37. A utilised GPU-hour costs about US$22 at base-case demand, against a market rate of US$6.71 today and roughly US$4.50 averaged over the hardware's life. The base case thus implies a subsidy of about US$61 million a year.
38. If the concern is about Ministers and their Directors uploading sensitive documents into Gemini or Kimi, then why not a simpler sovereign AI assistant for the civil service first?
39. That would need roughly 144 GPUs: about 7×10¹⁷ operations a second at 8-bit, US$13 million up front and US$6.7 million a year. That is under a thousandth of the 2027 target, and a far more realistic initial-term strategy. Does Ghana really need to train its base model when orchestrators now allow one to host their own data and switch across multiple advanced open-weight models to serve an Assistant or Agentic system?
40. Howzat?
Clif retweeted
Day 12 of looking for app Testers for hire
( You are getting paid to test apps)
Starting an "App testers for hire" channel for gh developers who want to hire testers for a fee, they will help test and find issues etc to help you pass testing phase. Those interested can join
@koboateng
@clifcode
t.me/+FW4Sz3xiYX1iMDRk
