Tristan Cunha retweeted
I pointed Claude at the below post and it embedded all 146,060 Steam games, sampled 320,000 points between them, ranked them by how far they land from anything that exists, and used creative judgment to pick a winner... and it cooked!
It's pick: Short Position, a co-op factory tycoon where one player has secretly shorted the stock.
4–8 players run one industrial empire: mines, smelters, rail, contracts. There’s no kill button and no sabotage minigame. The traitor plays the factory game badly on purpose. Overbuild a warehouse, route ore to the wrong smelter, sign a contract with a penalty clause on page 4. Every move looks like an honest mistake.
The twist is that suspicion is expensive. Anyone can audit the production logs, but audits cost money and spook investors. The traitor doesn’t need to wreck the factory. They just need everyone paranoid enough to do it themselves.
Each quarter ends with an earnings call and a board vote. Fire an honest exec and you lose their whole department. Four quarters, ~40 minutes.
It sits farther from its nearest neighbors than 83% of real Steam games. Factory games have co-op, social deduction games have tasks, and nobody has made the work itself the place where you lie.
It's pick: Short Position, a co-op factory tycoon where one player has secretly shorted the stock.
4–8 players run one industrial empire: mines, smelters, rail, contracts. There’s no kill button and no sabotage minigame. The traitor plays the factory game badly on purpose. Overbuild a warehouse, route ore to the wrong smelter, sign a contract with a penalty clause on page 4. Every move looks like an honest mistake.
The twist is that suspicion is expensive. Anyone can audit the production logs, but audits cost money and spook investors. The traitor doesn’t need to wreck the factory. They just need everyone paranoid enough to do it themselves.
Each quarter ends with an earnings call and a board vote. Fire an honest exec and you lose their whole department. Four quarters, ~40 minutes.
It sits farther from its nearest neighbors than 83% of real Steam games. Factory games have co-op, social deduction games have tasks, and nobody has made the work itself the place where you lie.
Of course it can. Take all 100,000+ games on steam, scrape their descriptions and metadata, semantically plot them in vector space using embeddings, mathematically choose a location in vector space far away from neighbors, convert that back to a description and metadata, boom, brand new ideas ranked by how unlike any other ideas they are.
48
28
879
57,797
If everyone was given a number based on web they were born, everyone could have a 4 digit sigil in base 1353.
Base-1353 encoding, so I can fit 256-bit values, like “96cca2a0c8a3263c74ef36274abf55f410d992034d16e8285811a9ad3fc59c13”, into short 26 character strings, like “ᓒᔳᣏዧᐎᑙᖜዌᙔᐸƫȶƃᒨዊጶᙵኻѫyፐᎉᑫᖫᘪᒩ”.
Also, I think these sigils are beautiful.
johnservil.github.io/sigil/play.html
0
0
1
30
Tristan Cunha retweeted
One of the most misinformed groups in the world has to be people who watch John Oliver religiously, this is one of the better forum posts on the phenomenon of the "comedic-news" segments
HBO's @LastWeekTonight has a long track record of citing social science findings without the appropriate context, massaging them to support whatever point @iamjohnoliver is trying to make.
125
681
8,747
453,944
Tristan Cunha retweeted
The true overall length of the Hawaiian archipelago is extremely surprising.
38
186
4,633
142,840
The ratios at the bottom of the chat are particularly impressive. The number of people living on $1/day dropped by 75%
It used to be that a quarter of everyone was living on just over $2 a day or less, now that's $5/day, over double.
It used to be that a quarter of everyone was living on just over $2 a day or less, now that's $5/day, over double.
No matter what poverty line you choose, the share of the global population below it has declined.
The World Bank's International Poverty Line is an extremely low poverty line. Set at $3 per day, it draws attention to the living conditions of the poorest people on the planet.
In n see how the share below this poverty line has declined over the last generation. In the year 2000, 36% of the world’s population lived on less than $3 per day. In 2024, it was 10%.
But is progress against global poverty only happening when measured against such a very low poverty line?
The chart gives the answer. The entire blue line is below the entire dark red line. This tells us that no matter what line you choose, the share below it has declined globally.
Progress against higher poverty lines has been large, too. The share living on less than $5 a day has declined from 57% to 25%, and the share living on less than $10 a day has declined from 74% to 53%.
But on the right-hand side of the chart, we see that most people still live on low incomes: 81% of people in the world live on less than $30 a day, or about $900 a month.
(written by @MaxCRoser)
0
0
0
21
Tristan Cunha retweeted
Look at what the Air Force Research Lab packed into its test pods on SpaceX's Starfall capsule.
Sensors, a satellite tracker, batteries, a water desalinator, survival gear and a medical kit. Engineers also checked that a life raft fits inside.
Every item on that list is for or cut off after a disaster.
It's still a research test. But a capsule that comes down from orbit with medicine and a way to make fresh water could reach places a truck or a ship can't get to for days.
Sensors, a satellite tracker, batteries, a water desalinator, survival gear and a medical kit. Engineers also checked that a life raft fits inside.
Every item on that list is for or cut off after a disaster.
It's still a research test. But a capsule that comes down from orbit with medicine and a way to make fresh water could reach places a truck or a ship can't get to for days.
17
61
309
15,310
Tristan Cunha retweeted
We dirty capitalists actually come up with a way of putting a value on people's lives.
Just ask them and they’ll tell you. "What would you be willing to pay for a 0.001% reduction in your individual risk of dying?" You can, also, infer it directly from behavior: from the workers make, the price-safety trade-offs customers make, etc.
This is called the Value of a Statistical Life (VSL).
I just asked Claude to run a regression of VSL against economic freedom (a proxy for capitalism, security of property rights, etc.).
And guess what? The more capitalist a country is, the more people value their lives there!
Just ask them and they’ll tell you. "What would you be willing to pay for a 0.001% reduction in your individual risk of dying?" You can, also, infer it directly from behavior: from the workers make, the price-safety trade-offs customers make, etc.
This is called the Value of a Statistical Life (VSL).
I just asked Claude to run a regression of VSL against economic freedom (a proxy for capitalism, security of property rights, etc.).
And guess what? The more capitalist a country is, the more people value their lives there!
8
34
338
5,490
Tristan Cunha retweeted
This is reality. The old Dutch landscape versus the new. This was an early morning last week on the Dutch countryside. While I am not a big fan of the massive turbines in the landscape, in this particular scene it looks somewhat unreal and spectacular.
17
141
1,384
17,022
Tristan Cunha retweeted
This is my methodology for finding huge performance gains in open-source software, primarily ffmpeg:
3 days ago I ran a full sweep of ffmpeg using dozens of Haiku threads to detect good candidates for easy optimizations, bug fixes and code cleanups (each thread searches for 1 of ew Sonnet threads go through every finding to prune out the false positives
Then I dig into each issue one by one, I don't rush things because ffmpeg is an extremely important library where reliability is super important
For each issue I take on, I take my sweet time with the help of Sonnet/Opus to implement the best possible solution and thoroughly test it every way I can
When I say thoroughly, I mean thoroughly lol
Yesterday to keep things simple one of my changes required changing a multiplication into a division, this change should theoretically be safe but to be 100.00% certain I had Sonnet design a test to go through every single one of the 4 billion possible values of floats
So now I know for 100.00% sure that the change works perfectly, since I've tested every possible value 😂
Then of course if the solution was crafted by Claude, I make sure to understand and question every line of code, hopefully this goes without saying
Essentially, the goal is to prove the solution wrong every way you can, even if it's a very obvious 1 line change that you're certain of
When you use LLMs for adversarial proof of your changes, it's ideal to run a new session that will not have any of the context of how the model arrived at that solution
So far I've gotten 0 pushback on any of the 24 PRs I sent and 16 have already been approved
The only comment I've received so far was a reviewer who asked me to implement a similar fix elsewhere in the file
If you have strategies to share to improve my work, let me know
3 days ago I ran a full sweep of ffmpeg using dozens of Haiku threads to detect good candidates for easy optimizations, bug fixes and code cleanups (each thread searches for 1 of ew Sonnet threads go through every finding to prune out the false positives
Then I dig into each issue one by one, I don't rush things because ffmpeg is an extremely important library where reliability is super important
For each issue I take on, I take my sweet time with the help of Sonnet/Opus to implement the best possible solution and thoroughly test it every way I can
When I say thoroughly, I mean thoroughly lol
Yesterday to keep things simple one of my changes required changing a multiplication into a division, this change should theoretically be safe but to be 100.00% certain I had Sonnet design a test to go through every single one of the 4 billion possible values of floats
So now I know for 100.00% sure that the change works perfectly, since I've tested every possible value 😂
Then of course if the solution was crafted by Claude, I make sure to understand and question every line of code, hopefully this goes without saying
Essentially, the goal is to prove the solution wrong every way you can, even if it's a very obvious 1 line change that you're certain of
When you use LLMs for adversarial proof of your changes, it's ideal to run a new session that will not have any of the context of how the model arrived at that solution
So far I've gotten 0 pushback on any of the 24 PRs I sent and 16 have already been approved
The only comment I've received so far was a reviewer who asked me to implement a similar fix elsewhere in the file
If you have strategies to share to improve my work, let me know
The optimization continues using Claude!
While my notifications filled up with criticism about me finding sloppy human code using Claude, I was busy optimizing one of the most widely used libraries in the world (ffmpeg)
This time, thanks to Claude's discovery, anybody who draws do so 8-24% faster
The 95-98% saving in dblur that I tweeted about yesterday was also merged into the project today
3 additional change suggestions of mine have been merged in the last 24 hours, all minor performance improvements related to loops performing an operation that needed to be performed only once
27
19
389
14,816
Tristan Cunha retweeted
I cannot let a steam engine analogy go by without bringing up John Wilkinson, perhaps the most underappreciated engineer of the Industrial Revolution.
Watt had figured out how to dramatically improve the steam engine, but he spent years struggling to actually manufacture one. The problem was embarrassingly simple: nobody could make a sufficiently round cylinder.
The casting and machining tolerances were atrocious. Watt complained that one 18-inch cylinder was out of round by as much as THREE-EIGHTHS OF AN INCH. Steam leaked past the piston. He tried packing the gaps with cork, oiled rags, paper, and apparently even old hats. Nothing worked particularly well.
Enter John Wilkinson, a British ironmaster whose immediate concern was manufacturing better cannons.
At the time, iron cannons were commonly cast around a core, introducing defects and misalignment that could make the barrels inaccurate or dangerously weak. Wilkinson developed a process for boring cannons from solid castings, and in 1774 patented a vastly improved boring machine.
The key innovation was almost comically mechanical: a massive, rigid boring bar, supported at both ends, that could cut a genuinely straight, circular hole instead of following the irregularities of the casting.
He adapted this technology to Watt's cylinders. By 1776, Matthew Boulton was boasting that Wilkinson had bored a FIFTY-INCH cylinder accurate to within the thickness of an old shilling.
Suddenly, the steam engine was a commercially viable machine.
And the story gets better. Wilkinson became one of Watt's first customers, buying an engine to drive the blast machinery at his own ironworks. The steam engine was now helping manufacture the iron that made more steam engines possible.
I find this whole episode absolutely extraordinary. The Industrial Revolution was, for a critical period, bottlenecked on the ability to make a sufficiently round hole.
We remember Watt, and rightly so. We remember the invention of the steam engine and the subsequent development of thermodynamics. But the man who made the machinery that made the whole thing possible is practically a footnote.
I think we consistently underestimate how much of scientific progress is actually downstream of advances in tooling, measurement, and manufacturing precision. It's not just that technology precedes theory. Sometimes the decisive breakthrough is that someone finally figures out how to build the thing the theorists have been talking about for twenty years.
Anyway, I think John Wilkinson deserves considerably more credit for modern civilization than he gets.
Watt had figured out how to dramatically improve the steam engine, but he spent years struggling to actually manufacture one. The problem was embarrassingly simple: nobody could make a sufficiently round cylinder.
The casting and machining tolerances were atrocious. Watt complained that one 18-inch cylinder was out of round by as much as THREE-EIGHTHS OF AN INCH. Steam leaked past the piston. He tried packing the gaps with cork, oiled rags, paper, and apparently even old hats. Nothing worked particularly well.
Enter John Wilkinson, a British ironmaster whose immediate concern was manufacturing better cannons.
At the time, iron cannons were commonly cast around a core, introducing defects and misalignment that could make the barrels inaccurate or dangerously weak. Wilkinson developed a process for boring cannons from solid castings, and in 1774 patented a vastly improved boring machine.
The key innovation was almost comically mechanical: a massive, rigid boring bar, supported at both ends, that could cut a genuinely straight, circular hole instead of following the irregularities of the casting.
He adapted this technology to Watt's cylinders. By 1776, Matthew Boulton was boasting that Wilkinson had bored a FIFTY-INCH cylinder accurate to within the thickness of an old shilling.
Suddenly, the steam engine was a commercially viable machine.
And the story gets better. Wilkinson became one of Watt's first customers, buying an engine to drive the blast machinery at his own ironworks. The steam engine was now helping manufacture the iron that made more steam engines possible.
I find this whole episode absolutely extraordinary. The Industrial Revolution was, for a critical period, bottlenecked on the ability to make a sufficiently round hole.
We remember Watt, and rightly so. We remember the invention of the steam engine and the subsequent development of thermodynamics. But the man who made the machinery that made the whole thing possible is practically a footnote.
I think we consistently underestimate how much of scientific progress is actually downstream of advances in tooling, measurement, and manufacturing precision. It's not just that technology precedes theory. Sometimes the decisive breakthrough is that someone finally figures out how to build the thing the theorists have been talking about for twenty years.
Anyway, I think John Wilkinson deserves considerably more credit for modern civilization than he gets.
a few thoughts:
1. pretraining science is kind of like how people built quite good steam engines before they understood thermodynamics. “Science owes more to the steam engine than the steam engine owes to science.”-(L.J. Henderson). if nick is right, which i am pretty sure he is, progress points in the direction of more interpretability, and pretraining was "slouching towards monosemanticity". my guess would be that major breakthroughs in interp also advance capabilities
2. nick is a good friend of mine, safe to say my origin story at openai wouldn't have happened if i hadn't met him, and it's fascinating watching him operate and seeing the early drafts evolve into this. he's the most profoundly happy guy I know- not driven by guilt, activism, or religious terror, but by fascination. he of course understands the stakes of understanding neural networks, but it doesn't seem like the core reason he does it: he just loves the neural nets. afaict he's posting this from some beautiful hideaway in tokyo. there's a lesson in there for all the +EV maxers. i feel very grateful to have met him
3. i do think it's crazy how few people across the world were staring at neural nets growing: most have moved to a different level of behaviorist analysis, thinking about reward functions and test scores and whatnot. it does not bode that well for the major model companies, oai included, that fundamental breakthroughs like this are coming from outside. we've become too much of a pressure cooker focusing on local incentives, studying narrow tracks under old paradigms, mostly doing engineering, rather than fundamental research, and implicitly hoping the next model will figure it out
67
393
3,660
101,178
Tristan Cunha retweeted
a few thoughts:
1. pretraining science is kind of like how people built quite good steam engines before they understood thermodynamics. “Science owes more to the steam engine than the steam engine owes to science.”-(L.J. Henderson). if nick is right, which i am pretty sure he is, progress points in the direction of more interpretability, and pretraining was "slouching towards monosemanticity". my guess would be that major breakthroughs in interp also advance capabilities
2. nick is a good friend of mine, safe to say my origin story at openai wouldn't have happened if i hadn't met him, and it's fascinating watching him operate and seeing the early drafts evolve into this. he's the most profoundly happy guy I know- not driven by guilt, activism, or religious terror, but by fascination. he of course understands the stakes of understanding neural networks, but it doesn't seem like the core reason he does it: he just loves the neural nets. afaict he's posting this from some beautiful hideaway in tokyo. there's a lesson in there for all the +EV maxers. i feel very grateful to have met him
3. i do think it's crazy how few people across the world were staring at neural nets growing: most have moved to a different level of behaviorist analysis, thinking about reward functions and test scores and whatnot. it does not bode that well for the major model companies, oai included, that fundamental breakthroughs like this are coming from outside. we've become too much of a pressure cooker focusing on local incentives, studying narrow tracks under old paradigms, mostly doing engineering, rather than fundamental research, and implicitly hoping the next model will figure it out
1. pretraining science is kind of like how people built quite good steam engines before they understood thermodynamics. “Science owes more to the steam engine than the steam engine owes to science.”-(L.J. Henderson). if nick is right, which i am pretty sure he is, progress points in the direction of more interpretability, and pretraining was "slouching towards monosemanticity". my guess would be that major breakthroughs in interp also advance capabilities
2. nick is a good friend of mine, safe to say my origin story at openai wouldn't have happened if i hadn't met him, and it's fascinating watching him operate and seeing the early drafts evolve into this. he's the most profoundly happy guy I know- not driven by guilt, activism, or religious terror, but by fascination. he of course understands the stakes of understanding neural networks, but it doesn't seem like the core reason he does it: he just loves the neural nets. afaict he's posting this from some beautiful hideaway in tokyo. there's a lesson in there for all the +EV maxers. i feel very grateful to have met him
3. i do think it's crazy how few people across the world were staring at neural nets growing: most have moved to a different level of behaviorist analysis, thinking about reward functions and test scores and whatnot. it does not bode that well for the major model companies, oai included, that fundamental breakthroughs like this are coming from outside. we've become too much of a pressure cooker focusing on local incentives, studying narrow tracks under old paradigms, mostly doing engineering, rather than fundamental research, and implicitly hoping the next model will figure it out
100
175
3,273
242,768
Tristan Cunha retweeted
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com
878
2,614
19,737
4,459,749
