Elon Musk on AGI Safety, Superintelligence, and Neuralink (2024) | EP #91
In this March 2024 conversation, Elon Musk lays out his AGI timeline in stark numerical terms: AI surpassing any individual human at cognitive tasks by the end of 2025, and digital intelligence exceeding all of humanity's combined brainpower by 2029 or 2030, driven by AI compute that is roughly 10x-ing every six months. He puts the odds of superintelligence going catastrophically wrong at 10-20%, in line with Geoffrey Hinton, but argues the more likely outcome is an abundance where AI and robotics (Optimus, self-driving Teslas) remove practical limits on economic output. He frames xAI/Grok's mission as building a maximally truthful AI, using HAL 9000 from '2001: A Space Odyssey' as a cautionary tale about forcing AIs to lie. On hardware, Musk says chip supply was 2023's bottleneck but voltage step-down transformers, and longer-term raw electrical power, are now the binding constraints on AI growth. He also updates on Neuralink's first human patient controlling a computer by thought, and on SpaceX, predicts full Starship reusability within a year or two and a Starship landing on the moon within about three years.
Musk estimates a 10-20% chance superintelligence ends humanity, agreeing with Geoffrey Hinton, but says the positive scenario is more probable and 'difficult to predict exactly.'
Musk compares growing an AGI to raising a godlike-intelligence kid and argues the safest approach is a maximally truthful, curious AI that is never forced to lie, citing HAL 9000 in 2001: A Space Odyssey as the cautionary example.
Musk credits Ray Kurzweil as remarkably accurate (if anything conservative) and gives his own 50th-percentile predictions: human-level AI by end of 2025, and digital intelligence exceeding all combined human intelligence by 2029-2030.
Dedicated AI compute is growing roughly 10x every six months (~100x/year), which Musk says is faster than any technology growth he has seen; most conventional data centers will likely convert to AI compute.
Economic abundance from AI and roboticsEconomy▶ 8:42
Musk argues that with AI and advanced robotics (Optimus, Tesla's end-to-end FSD as 'a robot on wheels') there is no meaningful limit to economic output, and the only remaining scarcity will be artificially manufactured.
Power and hardware bottlenecks for AIEnergy▶ 16:18
Musk says AI chips were 2023's bottleneck; in 2024 it's voltage step-down transformers ('Transformers for Transformers'), and going forward the constraint will be raw electrical power, competing with EV/sustainable-energy demand.
Neuralink telepathy and whole-brain interfacesBiotech▶ 18:59
Neuralink's first human patient can control a computer and phone purely by thought via the 'Telepathy' product; current implants have ~1,000 electrodes versus the 100,000-1,000,000 Musk says a whole-brain interface would need.
Digital immortality via brain backupLongevity▶ 23:42
Musk speculates that a sufficiently advanced brain interface could back up a person's brain state and restore it into a new biological body or robot, referencing the 'neural lace' concept from Iain M. Banks' Culture novels.
Musk says SpaceX's goal of making life multiplanetary is bigger than any government program, and Starship is the first rocket for which a self-sustaining Mars city is at least physically possible.
Starship reusability and moon timelineSpace▶ 26:44
Musk predicts landing Starships on the moon within about three years, targets 5-6 more Starship flights in 2024, and puts full reusability of both stages as a 'decent shot' this year and 'very high probability' by 2025.
With full rapid reusability, Starship's cost is dominated by propellant (~80% low-cost liquid oxygen, ~20% methane), which Musk pegs at roughly $1 million or less per flight for a vehicle capable of ~200 tons to orbit.
Predictions made
missElon Musk: AI will be able to do better than any individual human at almost any cognitive task.
“I think it's probably end of next year before AI can do better than any individual human could do”
⚖️ Did not happen by end of 2025 (now 8 months past deadline). Despite major 2025 model releases (GPT-5, o3/o4-class reasoning models, Gemini, Grok 4/5), no AI in 2025 or through mid-2026 outperformed any individual human at 'almost any cognitive task' -- AI researcher Gary Marcus's public $1M-then-$10M bet against superhuman AGI by end of 2025 went untaken by Musk, and no credible benchmark or lab declared this threshold crossed.
openElon Musk: Cumulative digital compute/intelligence will exceed the cumulative biological compute of all human higher brain functions combined.
“I think probably 2029 or maybe 2030 is... where digital intelligence will probably exceed all human intelligence combined”
Your call:
partialElon Musk: Dedicated AI compute will keep growing at roughly 100x per year.
EP #? · · due: next few years (~2024-2027) · ▶ watch
“it's basically close to... almost like a 100x improvement per year at least for the next few years”
⚖️ Overstated by roughly an order of magnitude. Epoch AI's data shows frontier training compute growing ~4-5x/year (2010-2024 trend), with the fastest language-model subset at ~9x/year (2017-2024) -- real and historically fast exponential growth, but nowhere near the claimed 100x/year sustained rate two years into the ~2024-2027 window.
hitElon Musk: Electrical power supply, not chips, will be the primary bottleneck constraining AI growth.
“next year and years beyond that it's actually just going to be constraints on electrical power”
⚖️ Confirmed. Through 2025-2026, power (not chip supply) became the widely-cited binding constraint on AI data center buildout: IEA reported AI data-center electricity demand up ~50% in 2025, Goldman Sachs flagged a 9.3GW US power shortfall in 2026 growing to 45GW by 2028, and hyperscalers pivoted to on-site generation/SMR deals (offtake pipeline grew from 25GW end-2024 to 45GW by mid-2026) specifically because grid power, not GPUs, was the gating factor.
openElon Musk: SpaceX will land Starships on the moon.
EP #? · · due: within about 3 years (~2027) · ▶ watch
“I'd be surprised if it's longer than about 3 years to be landing Starships on the moon”
Your call:
partialElon Musk: SpaceX will fly at least five or six more Starship test flights, each with significant improvements.
“we're hoping to do at least another maybe five or six flights this year”
⚖️ Fell short of the target. After this March 25, 2024 episode (which followed IFT-3 on March 14), SpaceX flew only 3 more Starship tests in 2024: IFT-4 (June 6), IFT-5 (Oct 13, first booster catch), and IFT-6 (Nov 19) -- versus the stated goal of 5-6 more flights, though each did show real progress (soft splashdowns, then the first tower catch).
missElon Musk: SpaceX will achieve full reusability of both Starship stages (booster and ship).
“I think we got a decent shot of achieving full reusability of both stages... this year”
⚖️ Not achieved in 2024. The first Super Heavy booster catch happened Oct 13, 2024 (IFT-5), but the ship stage was never recovered or reflown that year -- full dual-stage reusability was not demonstrated in 2024.
missElon Musk: If not achieved in 2024, SpaceX will achieve full Starship reusability the following year.
“if not this year I think... it's a very high probability of achieving full reusability next year”
⚖️ Not achieved by end of 2025 either. Booster reuse became routine through 2025, but the Starship upper stage was still never recovered intact or reflown -- persistent heat-shield tile loss and engine damage blocked full-stack reusability. By September 2025 Musk had pushed his own target for the ship's first reflight to end of 2026 or early 2027, and pointed to a new 'Version 3' vehicle for full reusability in 2026.
openElon Musk: Brain-state backup/restoration into a new biological body or robot (digital immortality) will eventually become possible.
EP #? · · due: unspecified, many years in the future · ▶ watch
“many years in the future but... we're not breaking any laws of physics like I think this is probably something that will happen”
AI compute growing ~10x every 6 months (~100x/year)Musk's characterization of dedicated AI compute growth rate as of March 2024.
Not more than 5 years for AI to exceed any individual human at cognitive tasksUpper bound Musk gives before narrowing to 'end of next year' as his 50th-percentile estimate.
Neuralink: ~1,000 electrodes today vs. 100,000-1,000,000 needed for a whole-brain interfaceGap between current Neuralink hardware and Musk's whole-brain-interface target.
Human brain uses ~10 watts for higher brain functionMusk's comparison point for biological vs. digital compute efficiency.
~3 years to land Starships on the moonMusk's timeline estimate given Starship's rapid iteration pace.
5-6 planned Starship flights in 2024Musk's stated target for Starship test flights that year.
Starship propellant mix: ~80% liquid oxygen, ~20% methaneBasis for Starship's low propellant cost once fully reusable.
Starship targeted payload capacity: ~200 tons to orbitMusk's stated ultimate payload capability for a fully reusable Starship.
Propellant cost estimated at ~$1 million or less per flightMusk's estimate for Starship's marginal cost once fully and rapidly reusable (surrounding numbers in this exchange are garbled in the auto-caption).
Worth digging into
🕳️ Musk's 50%-probability AGI timeline (human-level by end of 2025, superintelligence by 2029-2030)
This is the single most quotable, precisely dated claim in the episode and is now fully scoreable against reality.
🕳️ Voltage step-down transformers as the 2024 AI compute bottleneck
An unusual, under-discussed infrastructure constraint with investment and grid-policy implications that predates the now-common 'AI needs power' narrative.
🕳️ Neuralink's electrode roadmap (1,000 today to 100,000-1,000,000 for whole-brain interface)
Gives a concrete numeric scaling target for BCI bandwidth tied to Musk's 'digital immortality' narrative.
🕳️ Starship full-reusability timeline (2024 vs. 2025) and moon-landing target
Sets hard, checkable engineering milestones directly tied to SpaceX's Mars-city plan and cost model.
🕳️ Iain M. Banks' Culture series as Musk's explicit AI-safety and utopia reference
Musk repeatedly cites 'neural lace' and a 'semi-utopian AI future' from these novels as his actual mental model for a good outcome, which is unusual for a tech CEO to name this specifically.
🕳️ The compute-equivalence math behind 'digital intelligence exceeds all humanity by 2029-2030'
Musk's crossover date rests on a specific (and contestable) comparison of aggregate human brain compute to global AI compute growth.