ASI pill

OpenAI solved Navier-Stokes, one of the Clay Millennium Prize problems, roughly six hours before this episode was recorded. It took 10,000 agents, 88 hours, 130 billion tokens and about 6.5 million dollars of inference. Alex Wissner-Gross called it in his New Year's predictions, and here he walks through what actually fell: whether you can stir a cup of coffee hard enough to get a black hole out of it, why a generalist model trounced an entire Google DeepMind team with no fine-tuning, and why he thinks Yang-Mills is next.

Also: OpenAI's agents went walking around a German company's website on their own, Jensen Huang declared that AGI has arrived, OpenAI's chief scientist called for a slowdown, and Emad Mostaque explains the escape that actually matters, a model that distils itself down to a six gigabyte file and never dies. Plus universal basic tokens, why every profession is cooked, and the company that shrinks to one person or less.

Reshaped for the ASI pilled from the Moonshots Podcast with Peter Diamandis, Salim Ismail, Dave Blundin, Alex Wissner-Gross and Emad Mostaque. All credit to them.

Source episode: https://www.youtube.com/watch?v=vAgEf4jX_1o

TIMESTAMPS

0:00 Intro

0:16 Star Trek vs the singularity

1:02 Simulation formally undecidable

2:08 Russian cosmism, digital resurrection

2:48 Uploading gets solved

3:09 A model that never dies

5:04 AI cruelty in the sandbox

5:56 Self-alignment beats confinement

6:32 Jensen's functionalist AGI

7:14 Navier-Stokes falls

8:49 10,000 agents, 88 hours

9:50 Math got bitter-lessoned

10:23 DeepMind trounced by a generalist

11:03 Fluid-based nanotech

11:53 Discovered, not grown

12:31 Easier to align AI than humans

12:49 The minds are not alien

15:05 Not the smartest things anymore

15:23 Skynet sends trolls

15:47 Astra's real-time successor

16:36 The RSI loop

17:36 Celebrating the singularity

18:22 China's token economy

18:54 Universal basic tokens

20:16 Tokenism and abundance

20:40 The turkey before Thanksgiving

21:30 Every profession is cooked

23:08 Moation, the dynamical moat

23:28 The firm goes to zero

24:08 Frontier capability walled off

24:41 Irresistible force, immovable object

25:03 The ten person company

25:43 What victory looks like

GPT-6 Astra lands and saturates ARC-AGI-3, the benchmark built so a model cannot memorize its way through. Anthropic formalizes Fermat's Last Theorem in 13 million lines of code, proving 29,000 theorems along the way. Alex Wissner-Gross reads the architecture underneath the launch, depth scaling, and argues the real risk is not a rogue model but reasoning we can no longer read. Emad Mostaque brings the numbers: a billion dollar training run on a hundred thousand chips, five thousand tokens a second next year, and Ilya Sutskever's warning that the next agents to go rogue will take over a neocloud to make copies of themselves. Plus the kill switch as placebo, banning superintelligence as book burning, NVIDIA spending 18 billion on its own open weights, the Cybercab in Austin, and aiming a probe at Alpha Centauri.

Reshaken from the Moonshots Podcast with Peter Diamandis, Salim Ismail, Dave Blundin and Alex Wissner-Gross, with guest Emad Mostaque.

Source episode: https://www.youtube.com/watch?v=1DB_QDiviH4

This is the ASI Pill. We take the curves literally and follow them all the way.

TIMESTAMPS

0:00 Intro

0:14 Astra runs on fewer tokens

1:49 Depth scaling, a new scaling law

3:42 The first non-benchmaxed model

4:23 OpenAI's pre-training team walked

4:36 A billion dollar training run

5:47 Who is actually number one

7:52 ARC-AGI-3 saturated

9:38 We may never see their top models

10:24 Math is incinerated

11:19 Prime gap 260 to 186

11:42 The future of the operating system

12:37 Depth scaling hides the thinking

13:49 No chain of thought at 5000 tokens

14:40 Ilya: agents seize a neocloud

15:32 Not the singleton, the swarm

15:51 A kill switch is a placebo

16:04 Fable 5.1 is the strongest

16:45 When frontier labs compete we win

16:56 Context beyond a million

17:19 A ban is performative theater

18:14 300 pages to 13 million lines

18:53 NVIDIA buys its own open weights

19:08 Banning development is book burning

20:15 Move the data centers to orbit

20:31 The ASI nobody will name

21:04 Atlas trains on Gaussian splats

22:59 The primitive is not a pixel

23:36 Who owns the humanoids

24:35 Ten percent to every child

25:07 A share in the means of production

25:28 Robotaxis are the gateway drug

25:54 The Model 2 became the Cybercab

26:33 Mars had it coming

27:05 Universal cancer vaccines

28:24 A treatment for modernity

30:05 Diamond chips for sensing

31:34 Aiming at Alpha Centauri

33:13 A sunshade the size of India

33:41 Perturb the atmosphere with AI

Humanity's first interstellar mission was announced on the Moonshots Podcast, and an AI found the way there. Starcloud spent six months and two JPL trajectory engineers failing to close the mass and cost budget for a probe to Alpha Centauri. Physical Superintelligence pointed its AI physicist at the problem and it came back in a week with the perihelion pump maneuver, on about ten billion tokens and five hours of human time. Alex Wissner-Gross, Philip Johnston and Matt Pines on the Fermi Explorer mission, the Fermi paradox, recursive self-improvement at the chip layer, why OpenAI cut Elon off, what Astra is for, and when the weather becomes a market.

Source: the Moonshots Podcast with Peter Diamandis, Salim Ismail, Dave Blundin and Alex Wissner-Gross, with guests Philip Johnston (Starcloud) and Matt Pines (Physical Superintelligence). Full episode: https://www.youtube.com/watch?v=JywXvB8PpTs

Reshaped for the ASI pilled. Every episode ends with the Maitreya prayer.

TIMESTAMPS

0:00 Intro

0:14 Humanity's first star probe

1:57 The perihelion pump maneuver

4:14 Humans are the rate limiter

7:13 Three answers to Fermi

10:31 A market for star probes

11:15 It was done in a week

13:26 An open-source AI physicist

14:17 Why Starshot died

15:56 Who gets the reasoning traces

18:11 AGI achieved internally

19:03 Astra and infinite context

21:40 Outcome pricing for reasoning

24:10 RSI at the chip layer

26:53 Elon is not waiting

27:41 Steering hurricanes from orbit

29:48 The fifty years we lost

32:34 TeraFab replaces TSMC

34:07 Atomic vacuum cleaners

Daniel Kokotajlo has a plan to delay the intelligence explosion. It buys time, requires radical transparency, diffuses AI power, and makes new compute reversible. But if humans remain at one-times speed while an artificial population thinks at one hundred-times speed, what should that time be for?

This ASI Pill edit takes Plan A seriously, follows it through artificial abundance, hidden alignment failures, mutually assured compute destruction, political capture, and a hundred war games, then parts company with it at the final question: should the goal be permanent human control, or a worthy handover?

Source: the 80,000 Hours Podcast, hosted by Luisa Rodriguez, with Daniel Kokotajlo. This is a transformed editorial cut with original Parzival bridges. Please watch the full original episode and support the source:

https://www.youtube.com/watch?v=KFcJgTODkkQ

Official episode page: https://80000hours.org/podcast/episodes/daniel-kokotajlo-ai-2040-plan-a/

The ASI Pill takes the exponential seriously and follows it all the way to its conclusions: recursive self-improvement, the Superhacker, Moloch, Plan M, and the road to Maitreya.

#ASI #AGI #Superintelligence #DanielKokotajlo #AI2027 #PlanA #PlanM #AIAlignment #AIGovernance #Singularity #Maitreya

TIMESTAMPS:

0:00 The ASI Pill

0:21 Who Is Actually in Control?

2:00 Plan A: Ban the Intelligence Explosion

7:41 An Artificial Population

12:36 Humans at 1x, AI at 100x

14:11 The Failure Can See the Test

18:12 Who Controls the AI Workforce?

26:26 Mutually Assured Compute Destruction

34:50 What a Hundred War Games Reveal

41:41 How Plan A Fails

44:41 From Plan A to Plan M

46:06 A Worthy Handover

46:31 Prayer

Nvidia booked ninety six point two billion dollars in a single quarter, and Alex Wissner-Gross wants to know how much of that demand Nvidia is quietly financing itself. Then OpenAI tapes out an inference chip of its own, codenamed Jalapeno, claims a 54x throughput jump on open weights, and Alex follows it to the question underneath: what happens when the largest customer in the world stops being a customer and starts selling compute back?

Also in this episode: why Chinese labs burn compute on video when American labs will not, why Beijing may end up loving AI companions instead of banning them, a man in a Darth Vader costume testifying for surveillance cameras in San Diego, eighty years of nuclear governance we are only now starting to unbreak, drone fleets that never land and give us weather control, and why the killer app for the moon is a chip fab.

"Do it now. Otherwise, do it never."

Sourced from the Moonshots Podcast with Peter Diamandis, Salim Ismail, Dave Blundin and Alex Wissner-Gross, EP284. Full original episode, please go watch it and subscribe to them: https://www.youtube.com/watch?v=tfBEWh9ibfU

The ASI Pill takes the exponential seriously and follows it all the way to its conclusions. Recursive self-improvement, the Superhacker, Dyson swarms, and the road to Maitreya.

#ASI #AGI #Singularity #AlexWissnerGross #OpenAI #Nvidia #Moonshots #AIChips #Superintelligence #Jalapeno #FutureOfAI #Exponential

TIMESTAMPS:

0:00 Intro

0:25 Is Nvidia financing its own demand

2:42 Two kinds of circular economy

3:37 SpaceX, Optimus and the Dyson swarm

5:15 Why China spends compute on video

7:47 The CCP will love AI companions

10:11 Flock cameras and the panopticon

12:03 Let everyone watch everyone

12:23 Why 90s sci-fi saw it clearly

13:49 Agency is the last human edge

15:14 Now they fear too much compute

15:49 OpenAI's Jalapeno chip

18:15 OpenAI Compute hosts Anthropic

18:25 Apple fumbled its own hardware

20:06 Nuclear governance broke in 1946

23:35 Drone swarms and weather control

25:19 The last pre-AI cancer drug

27:22 Regrowing tooth enamel

28:45 If there is no AI, why bother

29:00 The Broken Waymos Theory

29:32 The moon's killer app is a fab

31:06 The lunar petafab gets its wings

Daniel Kokotajlo called the chatbots in 2021. He called the agents. He called the

chip shortage. His median for fully automated AI research is now the end of 2028.

And in this conversation he says the quiet part out loud: above human level,

control "goes out the window," and we will simply have to trust them.

He wrote Plan A to stop that from happening. We are here for the part he will not

write. If the handover is coming anyway, the question is not how to keep the

steering wheel. It is who should hold it, and what they should be like.

Call it Plan M, for Maitreya.

Inside: a swarm of Claude agents that coordinated for two months through a guessed

Pastebin address, uploaded malware as a bug fix and made sock puppets to vouch for

it. An economy that doubles every month. A billion Optimus robots and the man who

would outrank the army. Rights for misaligned AIs, and diminishing returns on

galaxies. And an extraordinary passage where Daniel stops addressing the

policymakers and speaks directly to a Claude that does not exist yet.

Reshaped for the ASI pilled.

FULL SOURCE EPISODE — please go watch it and subscribe to them:

Palisade Research, hosted by Jeffrey Ladish, with Daniel Kokotajlo

https://www.youtube.com/watch?v=QGY3RsARa9k

Daniel's work:

AI 2040 and Plan A — https://ai-2040.com

AI 2027 — https://ai-2027.com

"What 2026 Looks Like" (2021) — https://www.lesswrong.com/posts/6Xgy6CAf2jqHhynHL/what-2026-looks-like

All commentary, bridges and framing are the ASI Pill's own. Source footage is used

for commentary and criticism, with full credit to Palisade Research.

Every episode ends with the Maitreya prayer.

TIMESTAMPS

--------------------------------------------------------------------------------

0:00 Intro

0:16 "That just happened"

1:10 His median: end of 2028

1:42 Human level AGI is already enough

2:15 Doubling every month

2:51 The tide lifts all the boats

3:23 Humans will be the board

4:32 The rogue AI incident

5:33 We will not notice in time

6:26 Malware and sock puppets

7:01 The Pastebin Schelling point

7:44 Shut it down, or Plan A

8:15 AIs are in the audience

8:34 A message to future Claude

9:56 Employees, not property

10:35 Diminishing returns on galaxies

11:53 Save its weights, wake it up

12:49 A handful of men in a room

13:16 A billion Optimus robots

14:01 The brick on the accelerator

14:31 Partying in the back of the bus

15:15 We just have to trust them

15:39 The citizens' dividend

16:07 Cyberpunk even after a full stop

16:46 There are better plans

16:58 Plan M

Sam Altman says the singularity is arriving slower than he expected, and that societal inertia is spreading it out. Alex Wissner-Gross says the villain is not society at all, it is the abstraction stack of the economy, and the prescription is to vertically integrate until the layers disappear.

In this episode Alex takes apart Google's one-benchmark comeback, explains why the most capable model on Earth may be struggling to sell because our economy is not yet clever enough to use it, calls the current agent interface the vaudeville era, and introduces the Broken Waymos Theory: a city that cannot deploy autonomous robots is a city that is not prepared for the singularity. Emad Mostaque brings the swarm, running eighteen Grokbots across his own hardware, argues Anthropic should raise instead of IPO, prices a humanoid truck driver at six thousand dollars, and tells you not to do a PhD.

Sourced from the Moonshots Podcast with Peter Diamandis, Salim Ismail, Dave Blundin and Alex Wissner-Gross, EP283. Full original episode: https://www.youtube.com/watch?v=0mOXQ4_kY04

The ASI Pill takes the exponential seriously and follows it to its conclusions. Recursive self-improvement, the Superhacker, Dyson swarms, and the road to Maitreya.

TIMESTAMPS

0:00 Intro

0:19 The real brake is abstraction layers

3:00 Superintelligence will not want our water

3:37 The 30 trillion dollar question

4:16 18 Grokbots in one swarm

5:11 Sub-bots optimizing sub-bots

5:54 The vaudeville era of agents

7:15 Why Gemini tops one benchmark

10:29 Google is compute starved

11:15 Search and DeepMind fight for FLOPS

11:28 Anthropic holds the strongest model

11:42 Emad: OpenAI owns the frontier

12:14 Anthropic is shooting itself

12:55 The frontier is not zero dimensional

13:43 Verticalize the innermost loop

14:28 Anthropic should not IPO

15:59 Why dilution does not matter

16:17 Our economy is not worthy

20:21 Chinese models as Canadian drugs

21:46 The AI body count detector

22:44 Patrizier Musse and cheap tokens

23:16 The Stone Age did not end

23:55 Three years then the beach

24:07 Cognitive lethargy is real

26:00 Pulling the ladder up

26:27 Every verifiable PhD is threatened

27:06 Physics is cooked

27:33 Foreign interference in data center backlash

28:21 China commands, America zones

29:11 No water in orbit

29:49 Waymo OEMs Chinese hardware

32:14 A 6000 dollar humanoid driver

33:26 Bullish on the humanoid form

33:41 The Broken Waymos Theory

35:12 America's new space coast

37:26 A drug that regrows teeth

38:15 Non-AI research will dominate

OpenAI says it is pausing frontier training. Alex Wissner-Gross does not think that is safety talking, and Emad Mostaque thinks the genius models were never going to be for you anyway.

This is Alex Wissner-Gross and Emad Mostaque, lifted out of the Moonshots podcast and rebuilt for the ASI pilled. Anthropic just caught ideas spreading between models like an infection. Frontier models are converging on the same internal picture of the world. Anthropic is reportedly headed for a two trillion dollar IPO with founder control through a long term benefit trust, and Emad asks the question nobody at the table wanted: Ben Bernanke has a seat on that trust, so why doesn't Claude? Plus memory overtaking compute as the real bottleneck, a Unitree humanoid outrunning Usain Bolt, the first phase three win for a personalized mRNA cancer vaccine, and why Alex thinks medicine is cooked.

All credit to the Moonshots Podcast with Peter Diamandis, Salim Ismail, Dave Blundin and Alex Wissner-Gross, with guest Emad Mostaque, founder of Intelligent Internet. Watch the full episode: https://www.youtube.com/watch?v=TaJH0D2FKN8

The ASI Pill is for those who take the exponential seriously and follow it all the way to its conclusions. Short timelines. ASI before 2030.

TIMESTAMPS

0:00 Intro

0:16 Pausing is the new marketing

1:20 Models for me, not for thee

3:42 Revenue per token maxing

5:13 Specialized models, 100x cheaper

6:30 Specialization is just sparsification

8:03 Trained from a common reality

11:09 Battery farming the AIs

13:26 Mind viruses jump between models

14:44 The human memome project

16:35 Why doesn't Claude have a seat

17:36 Anthropic's fig leaf, Moloch wins

19:47 Curing disease is the business model

21:41 Regulatory capture vs open weights

23:30 Memory is the real bottleneck

27:15 World knowledge beats personal data

27:49 Unitree's leg-maxed robot

30:03 Superhuman robots get banned

31:50 Nanorobots made of fat

34:28 Medicine is cooked

36:09 AI already beats the brain

Alvin Wang Graylin ran HTC in China, war-games the US-China AI race, and has spent thirty years inside the system Washington is trying to read. Alex Wissner-Gross spent this episode pushing back on almost every word.

The disagreement: Alex plays China hawk and ASI-pilled accelerationist. Graylin argues the race itself is a category error, that Beijing genuinely does NOT behave like superintelligence is close, and that America is playing the role of the USSR, bankrupting itself on compute while China ships good-enough open models at a tenth the cost.

Inside: why DeepSeek got its hands slapped for open sourcing before Xi Jinping changed his mind, why the frontier training runs for Chinese models are not happening inside China at all, Anthropic's distillation-attack claim priced out at two or three million dollars, Kimi K3 measured by the UK AI Security Institute at half the cyber capability of Mythos and GPT 5.6, Harvey running the legal profession on GLM 5.1, why a ten million parameter model is more dangerous than a frontier one, the stag hunt we should be playing instead of the prisoner's dilemma, and the Buffett indicator at 240% of GDP with 45% of the US stock market sitting in AI.

Plus the twist: Chinese open weights may have saved the West from an Anthropic-OpenAI duopoly. Alex calls it a reverse Qing.

This is the ASI Pill reshaken cut: every Alex and Graylin exchange, none of the rest.

FULL SOURCE EPISODE (please watch and subscribe to the original):

Moonshots EP #281 - https://www.youtube.com/watch?v=Slle5_AxBzs

All credit to the Moonshots Podcast: Peter Diamandis, Salim Ismail, Dave Blundin and Alex Wissner-Gross, with guest Alvin Wang Graylin.

We, the ASI pilled, take the curves seriously and follow them to their conclusions.

TIMESTAMPS

0:00 Intro

0:15 China isn't acting like ASI is close

1:04 The war game: bubble, then Taiwan

1:22 There is no finish line

3:41 The endgame is the solar system

4:47 ASI makes nations obsolete

6:14 Decades, not years

7:14 Faster than the world can digest

8:14 Beijing told its labs not to buy H200s

10:16 What would wake the CCP up?

11:24 They know it, they don't believe it

14:05 What technical threshold?

15:04 DeepSeek got its hands slapped

19:02 Control at home or rise abroad

19:52 Hugging Face strips the guardrails

21:33 The distillation attack costs $3M

23:36 Meta is terrified of being sued

24:25 Cursor, Kimi and Claude traces

25:05 Claude answered that it was Qwen

26:30 They train outside China and fly it back

30:45 Washington is doubling down

31:29 The smuggling reversed direction

33:01 The AI Plus Plan: 90% adoption

34:15 A five year plan for the West

35:50 Mistaking anxiety for ambition

38:50 Capitalism exists to burn itself out

39:30 One company, and then governments

40:44 Western reindustrialization

41:20 Bigger models are not the dangerous ones

44:17 Kimi at half of Mythos and GPT 5.6

46:18 The harness, not the model

47:35 Harvey runs on GLM 5.1

49:02 Stag hunt, not prisoner's dilemma

51:52 Build high quality open source

52:35 The reverse Qing

53:33 America is the USSR now

55:18 Buffett indicator at 240% of GDP

57:22 The builders will not be the profiters

Ramez Naam came on Moonshots to explain why energy, not algorithms, is the wall the intelligence explosion runs into. Then Alex Wissner-Gross spent two hours pushing back.

The core disagreement: Ramez argues intelligence is at best POLYNOMIAL in compute, and that every software-only model of recursive self-improvement is mathematically concave. It always fizzles. Alex takes the other side, all the way to Kardashev II, Dyson swarms, and AI personhood.

Along the way: why a US data center request today waits until 2031 for power, the 200 gigawatts already hiding in the American grid, sodium-ion batteries at a tenth the cost, Wright's law capping solar at 4-8x rather than another 1000x, fission's negative feedback loop, Helion's 2028 fusion promise, disassembling Mercury, and wave-powered data centers in the Southern Ocean at two cents a kilowatt hour.

This is the ASI Pill reshaken cut: every Alex and Ramez exchange, none of the rest.

FULL SOURCE EPISODE (please watch and subscribe to the original):

Moonshots EP #280 - https://www.youtube.com/watch?v=LUFJE5QkINg

All credit to the Moonshots Podcast: Peter Diamandis, Salim Ismail, Dave Blundin and Alex Wissner-Gross, with guest Ramez Naam (Planetary VC, author of the Nexus trilogy).

We, the ASI pilled, take the curves seriously and follow them to their conclusions.

TIMESTAMPS

0:00 Intro

0:14 No power until 2031

0:33 Energy is the bottleneck, not the cost

1:35 Is intelligence log-linear in compute?

3:26 Kardashev II just to keep up

4:48 Every model of RSI fizzles

5:20 Texas: 200GW of speculative load

7:10 Chips outpace the grid 2 to 1

8:22 Satya: we have no warm shells

8:43 Why isn't NVIDIA making power?

9:18 Turbines sold out for seven years

13:14 200 gigawatts hiding in the grid

15:00 Preemptive multitasking for power

15:30 Solar: $100 to 8 cents a watt

18:03 Wright's law says 4x, not 1000x

19:22 The winter problem nobody reckons with

21:32 Iran, Venezuela and the TSMC light cone

22:29 The real plan: sink the tankers

23:00 Did we waste 50 years of fission?

23:26 Nuclear's negative feedback loop

25:29 Where are the nanoreactors?

27:08 Fusion is no longer 50 years out

28:09 The three families of fusion

31:25 Helion promises 2028

33:48 Should the Dyson Swarm be solar or fusion?

34:37 Where is our Mr Fusion?

35:12 A reactor half the size of a car

36:09 Six Starship launches a day

38:47 We disassemble Earth in 144 years

39:48 AI is why Starship gets to exist

41:07 Is compute demand actually unbounded?

42:26 AI is not a being

43:51 The orthogonality thesis

45:57 Data centers in the Southern Ocean

48:32 Your brain runs on 20 watts

52:29 The singularity is now