Garrett Baker — LessWrong

I have signed no contracts or agreements whose existence I cannot mention.

They thought they found in numbers, more than in fire, earth, or water, many resemblances to things which are and become; thus such and such an attribute of numbers is justice, another is soul and mind, another is opportunity, and so on; and again they saw in numbers the attributes and ratios of the musical scales. Since, then, all other things seemed in their whole nature to be assimilated to numbers, while numbers seemed to be the first things in the whole of nature, they supposed the elements of numbers to be the elements of all things, and the whole heaven to be a musical scale and a number.

Metaph. A. 5, 985 b 27–986 a 2.

I have signed no contracts or agreements whose existence I cannot mention.

They thought they found in numbers, more than in fire, earth, or water, many resemblances to things which are and become; thus such and such an attribute of numbers is justice, another is soul and mind, another is opportunity, and so on; and again they saw in numbers the attributes and ratios of the musical scales. Since, then, all other things seemed in their whole nature to be assimilated to numbers, while numbers seemed to be the first things in the whole of nature, they supposed the elements of numbers to be the elements of all things, and the whole heaven to be a musical scale and a number.

Metaph. A. 5, 985 b 27–986 a 2.

For Elon motivating SpaceX employees is important, so he needs to tell the public a story about how SpaceX isn't just about building AI datacenters even if he thinks building superintelligence is the main goal of SpaceX and it's important to build AI datacenters for that and he can't build enough of them on earth.

In the interview he argues that the reason he wants to build a ton of solar panels to capture a bunch of the sun's energy is because of AI, he really comes across as extremely AI pilled, he is not downplaying his exciting singularity-inducing AI-related objectives in order to convince SpaceX employees to work on making solar panels from moon dust.

As far as the AI datacenters in orbit go, it's not clear to me that the thermodynamics work out for that project and it's easy enough to lose the heat that the datacenters produce.

What thermodynamics are you concerned about? It is extremely efficient to radiate heat in space (a good thing too because that is your only option!). The equation here is the Stefam-Boltzmann law, which says

where $P$ is the power you radiate away, $ε$ is your emissivity, $σ$ is the Stefan-Boltzmann constant, $A$ is your surface area, and $T$ is your temperature. Note that that is $T^{4}$ , small changes in your temperature have large bang-for-your-buck in your radiation power. Intuitively, this should make things easy.

I got Claude to run the numbers, and Claude found that the launch cost of the radiator would only add $3M, assuming $200/kg. Meanwhile AI "hyperscaling" datacenters cost around $100M to build, and (assuming solar panels are already up there) weigh 200 tons leading to a launch cost of $40M. That is to say, compared to the other costs involved, the radiator seems very minor.

Claude's radiation analysis

Alright, let's just do vanilla flat-panel radiators at elevated chip temps.

Assumptions

Data center thermal load: $P = 10 MW$ (moderate — a hyperscaler pod is 20-50 MW, but let's start here)
Emissivity: $ε = 0.92$ (anodized aluminum or similar high-ε coating)
Two-sided radiation (both faces of panel radiate to space)
Background sink temperature: $T_{sink} \approx 3 K$ (negligible, so we ignore it)
Sun-shielded radiators (no solar absorption)

Stefan-Boltzmann

q = ε σ T^{4}

where $σ = 5.67 \times 10^{- 8} {W m}^{- 2} K^{- 4}$ .

Per side, per m²:

Radiator Temp	$q$ (W/m²)	Two-sided (W/m²)
350 K (77°C)	783	1,566
400 K (127°C)	1,336	2,671
450 K (177°C)	2,139	4,278
500 K (227°C)	3,261	6,522

Required Area for 10 MW

A = \frac{P}{2 ε σ T^{4}}

Radiator Temp	Area (m²)	Equiv. square panel side
350 K	6,386	~80 m
400 K	3,744	~61 m
450 K	2,338	~48 m
500 K	1,533	~39 m

At 500 K the radiator is roughly the footprint of a large warehouse. Not crazy.

Mass Estimate

Aluminum honeycomb radiator panels run about 3–5 kg/m² with embedded heat pipes for thermal spreading. Call it 4 kg/m².

Radiator Temp	Area (m²)	Radiator mass (tonnes)
350 K	6,386	25.5
400 K	3,744	15.0
450 K	2,338	9.4
500 K	1,533	6.1

At current Starship-aspirational costs (~$200/kg to LEO), the 500 K case is about $1.2M just for radiator launch mass. Very manageable relative to the cost of the compute hardware itself.

The Chip Temperature Constraint

Here's where it gets real. You need a thermal gradient from junction → coolant → radiator. Typical chain:

T_{junction} = T_{radiator} + Δ T_{interface} + Δ T_{transport}

If $Δ T_{transport} \approx 20 - 50 K$ (pumped liquid loop) and $Δ T_{interface} \approx 10 - 20 K$ , then:

Radiator Temp	Required $T_{junction}$
400 K	~450–470 K (177–197°C)
450 K	~500–520 K (227–247°C)
500 K	~550–570 K (277–297°C)

Standard silicon tops out around 150°C junction temp for reliability. SiC or GaN can handle 250°C+, but nobody's making data-center-class GPUs in those processes. So realistically you're stuck at ~400 K radiator temp unless you accept exotic semiconductors, which puts you at ~3,700 m² and 15 tonnes of radiator for 10 MW.

Bottom Line Spec Sheet (Conservative)

Thermal load: 10 MW
Radiator temp: 400 K
Radiator area: 3,744 m² (~61m × 61m)
Radiator mass: ~15 tonnes
Coolant: pumped single-phase liquid loop
Launch cost (radiators only): ~$3M at $200/kg
Constraint: must be sun-shielded or edge-on to sun

The $T^{4}$ scaling means even modest advances in high-temp compute (say, junction temps of 200°C becoming standard) buy you a lot of radiator mass savings.

Claude's datacenter launch cost analysis

Good question. Let me break this down by component. I'll need to look up some specifics on modern rack densities.Okay, great data. Let me work through this.

The Compute Hardware

The GB200 NVL72 rack weighs 1.36 metric tons and draws ~120 kW. But the full deployment arrives in four components: compute rack (1,500 kg), NVLink Switch rack (800 kg), CDU (400 kg), and PDU (300 kg) — totaling ~3,000 kg.

For 10 MW of IT load:

N_{racks} = \frac{10,000 kW}{120 kW/rack} \approx 83 racks

Mass Budget

Component	Per rack (kg)	× 83 racks (tonnes)	Needed in space?
Compute rack	1,500	124.5	Yes
NVLink switch rack	800	66.4	Yes
CDU (coolant dist.)	400	33.2	Redesigned¹
PDU	300	24.9	Redesigned²
Terrestrial total	3,000	249

¹ You still need pumped-loop coolant distribution, but the CDU on Earth interfaces with building chilled water. In space, it interfaces with the radiator loop directly — probably lighter.

² Power distribution from solar is DC→DC, not AC→DC. Simpler, lighter. Call it half mass.

Optimistic space mass estimate: strip CDUs to ~200 kg (simpler loop), PDUs to ~150 kg (DC-native), and you could probably shave ~20% off structural rack mass since you don't need to support against gravity (though you still need rigidity for launch loads and thermal expansion).

Component	Space-adapted (kg)	× 83 (tonnes)
Compute rack	1,200	99.6
NVLink switches	650	53.9
Coolant distribution	200	16.6
Power distribution	150	12.5
Space total	~2,200	~183

Add ~10% for structural housing, cabling, interconnects between racks, thermal interface hardware → ~200 tonnes for the compute payload.

Cost

Hardware procurement: All-in cost per GB200 NVL72 rack is about $3.9M for a typical hyperscaler, so:

83 \times $ 3.9 M \approx $ 324 M

Launch cost at $200/kg (aspirational Starship):

200,000 kg \times $ 200 / kg = $ 40 M

At current Falcon Heavy prices (~$2,000/kg):

200,000 kg \times $ 2,000 / kg = $ 400 M

Summary

Item	Cost
GPU/compute hardware	~$324M
Radiator hardware + fabrication	~$15M
Launch (compute, @$200/kg)	~$40M
Launch (radiators, 15t @$200/kg)	~$3M
Launch (compute, @$2000/kg)	~$400M
Total (optimistic Starship)	~$380M
Total (current launch prices)	~$740M

For reference, a comparable terrestrial datacenter is $100–150M. So you're paying a 3–5× premium at Starship prices, mostly for the hardware itself (which costs the same either way). The launch cost is actually not the dominant term at $200/kg — the GPUs are.

The real kicker: GPU lifetime is ~3 years before obsolescence. On Earth you swap them. In orbit, you've got a $324M asset that's depreciating at $108M/year and is very hard to service. That's arguably the actual killer for the economics, not the physics.

In his recent Dwarkesh Patel interview, Musk is pretty clear about why he wants to go to the moon: he wants to build a ton of solar panels to capture more of the sun's energy, its easier to launch from the moon than from the earth, and the moon has plenty of silicon ^[1] . He is also pretty clear that he wants to build datacenters in orbit. I don't think we need to speculate to get answers here.

Also, it keeps things interesting for the simulator gods. ↩︎

Wrong conclusions are inevitable and commonplace. Godel's Theorems apply to all formalisms.

A tangent, but Godel's incompleteness theorems simply show that for sufficiently powerful formal systems:

There are statements which are true, but unprovable
If the system is consistent, "this system is consistent" is such a statement.

Neither of which show that all formal systems are unsound. That is, if a statement is provable in a formal system, the corresponding property is true in all models of that formal system. So this point is not correct because of Godel (though it could be practically correct for other reasons, such as the world being complicated).

I don't know the reading list for the 17th yet, but will try to publish it on Wednesday, much sooner than typical, for your sake!

Here is the claimed Gorard's "alien logic system" in the linked tweet

I had Claude chew on this for a bit, and Claude determined that this proof system was the trivial one (ie ∀a,b:a=b) by writing a script in SPASS (an automatic theorem prover) to try to prove this statement.

Here is the SPASS proof

--------------------------SPASS-START-----------------------------
Input Problem:
1[0:Inp] || equal(skc2,skc3)** -> .
2[0:Inp] ||  -> equal(op(op(u,v),op(v,u)),v)**.
3[0:Inp] ||  -> equal(op(op(u,v),v),op(v,op(v,u)))**.
4[0:Inp] ||  -> equal(op(op(u,op(v,w)),w),op(v,op(u,op(w,u))))*.
 This is a unit equality problem.
 This is a problem that has, if any, a non-trivial domain model.
 The conjecture is ground.
 Axiom clauses: 3 Conjecture clauses: 1
 Inferences: IEqR=1 ISpR=1 ISpL=1 
 Reductions: RFRew=1 RBRew=1 RFMRR=1 RBMRR=1 RObv=1 RUnC=1 RTaut=1 RFSub=1 RBSub=1 
 Extras    : Input Saturation, No Selection, No Splitting, Full Reduction,  Ratio: 5, FuncWeight: 1, VarWeight: 1
 Precedence: op > skc3 > skc2 > skc1 > skc0
 Ordering  : KBO
Processed Problem:

Worked Off Clauses:

Usable Clauses:
1[0:Inp] || equal(skc3,skc2)** -> .
2[0:Inp] ||  -> equal(op(op(u,v),op(v,u)),v)**.
3[0:Inp] ||  -> equal(op(op(u,v),v),op(v,op(v,u)))**.
4[0:Inp] ||  -> equal(op(op(u,op(v,w)),w),op(v,op(u,op(w,u))))*.
	Given clause: 1[0:Inp] || equal(skc3,skc2)** -> .
	Given clause: 2[0:Inp] ||  -> equal(op(op(u,v),op(v,u)),v)**.
	Given clause: 3[0:Inp] ||  -> equal(op(op(u,v),v),op(v,op(v,u)))**.
	Given clause: 15[0:Rew:3.0,14.0,2.0,14.0] ||  -> equal(op(u,op(u,v)),op(v,op(v,u)))*.
	Given clause: 4[0:Inp] ||  -> equal(op(op(u,op(v,w)),w),op(v,op(u,op(w,u))))*.
	Given clause: 11[0:SpR:3.0,2.0] ||  -> equal(op(op(u,op(u,v)),op(u,op(v,u))),u)**.
	Given clause: 16[0:SpR:15.0,2.0] ||  -> equal(op(op(op(u,v),u),op(v,op(v,u))),u)**.
SPASS V 3.9 
SPASS beiseite: Proof found.
Problem: collapse.dfg 
SPASS derived 171 clauses, backtracked 0 clauses, performed 0 splits and kept 96 clauses.
SPASS allocated 85510 KBytes.
SPASS spent	0:00:00.08 on the problem.
		0:00:00.03 for the input.
		0:00:00.02 for the FLOTTER CNF translation.
		0:00:00.00 for inferences.
		0:00:00.00 for the backtracking.
		0:00:00.01 for the reduction.


Here is a proof with depth 3, length 23 :
1[0:Inp] || equal(skc3,skc2)** -> .
2[0:Inp] ||  -> equal(op(op(u,v),op(v,u)),v)**.
3[0:Inp] ||  -> equal(op(op(u,v),v),op(v,op(v,u)))**.
4[0:Inp] ||  -> equal(op(op(u,op(v,w)),w),op(v,op(u,op(w,u))))*.
14[0:SpR:2.0,3.0] ||  -> equal(op(op(u,v),op(op(u,v),op(v,u))),op(u,op(u,v)))**.
15[0:Rew:3.0,14.0,2.0,14.0] ||  -> equal(op(u,op(u,v)),op(v,op(v,u)))*.
16[0:SpR:15.0,2.0] ||  -> equal(op(op(op(u,v),u),op(v,op(v,u))),u)**.
17[0:SpR:15.0,2.0] ||  -> equal(op(op(u,op(u,v)),op(op(v,u),v)),op(v,u))**.
34[0:SpR:4.0,2.0] ||  -> equal(op(op(u,op(v,op(w,u))),op(w,op(v,op(u,v)))),op(v,op(w,u)))**.
49[0:SpR:4.0,2.0] ||  -> equal(op(op(u,op(op(op(v,u),u),v)),v),u)**.
63[0:Rew:3.0,49.0] ||  -> equal(op(op(u,op(op(u,op(u,v)),v)),v),u)**.
128[0:SpR:16.0,3.0] ||  -> equal(op(op(u,op(u,v)),op(op(u,op(u,v)),op(op(v,u),v))),op(v,op(u,op(u,v))))**.
145[0:SpR:3.0,16.0] ||  -> equal(op(op(op(u,op(u,v)),op(v,u)),op(u,op(u,op(v,u)))),op(v,u))**.
168[0:Rew:17.0,128.0] ||  -> equal(op(op(u,op(u,v)),op(v,u)),op(v,op(u,op(u,v))))**.
171[0:Rew:168.0,145.0] ||  -> equal(op(op(u,op(v,op(v,u))),op(v,op(v,op(u,v)))),op(u,v))**.
173[0:Rew:34.0,171.0] ||  -> equal(op(u,op(u,v)),op(v,u))**.
174[0:Rew:173.0,15.0] ||  -> equal(op(u,op(u,v)),op(u,v))**.
194[0:Rew:173.0,63.0] ||  -> equal(op(op(u,op(op(v,u),v)),v),u)**.
209[0:Rew:173.0,174.0] ||  -> equal(op(u,v),op(v,u))*.
228[0:Rew:173.0,194.0,209.0,194.0,173.0,194.0,173.0,194.0,209.0,194.0] ||  -> equal(op(u,v),u)**.
229[0:Rew:228.0,2.0] ||  -> equal(op(u,v),v)**.
234[0:Rew:228.0,229.0] ||  -> equal(u,v)*.
235[0:UnC:234.0,1.0] ||  -> .
Formulae used in the proof : conjecture0 axiom1 axiom0 axiom2

--------------------------SPASS-STOP------------------------------

Here is Claude's proof summary

Proof sketch (from SPASS, depth 3, length 23):

Axioms 1 & 2 ⟹ a ⊕ (a ⊕ b) = b ⊕ (b ⊕ a) [step 15: "symmetrized absorption"]
Substitutions into Axiom 3 + rewrites ⟹ a ⊕ (a ⊕ b) = b ⊕ a [step 173: "double apply = flip"]
(1) + (2) ⟹ a ⊕ (a ⊕ b) = a ⊕ b [step 174: idempotent-like]
(2) + (3) ⟹ a ⊕ b = b ⊕ a [step 209: commutativity]
(4) + earlier lemmas ⟹ a ⊕ b = a [step 228: left absorption]
Rewrite Axiom 2 with (5) ⟹ a ⊕ b = b [step 229: right absorption]
(5) + (6) ⟹ a = b ∀ a,b ∎

Therefore I think its quite likely that many of the supposedly rich alien axiom systems Gorard found are actually just trivial almost-contradictory systems which are hard to prove the triviality of. It also explains why he couldn't "make sense" of the system. There is nothing to make sense of.

I think the content and arguments in the possessed machines are not very good, the prose is ok but it would read as too self important and AI slop on a less distinguished webpage. I think that many people were charmed by how respectable and nice the website looked and so were willing to give the writing and arguments much more leeway when sharing the page (if they read it at all).

Edit: I believe this because I notice it is popular, and I noticed this dynamic in myself, I felt I wanted the article to be deep and interesting, but also that if I thought about telling the arguments to a skeptical observer very few of them would stand up to scrutiny.

Yeah, the possessed machines is a very good example of how important good web design is.

If you are to make claims like this, at least make arguments. This isn't twitter.

Caro is extremely comprehensive and will write small mini-books on the history of every significant institution or person LBJ ever touched. That means that The Master of the Senate begins in like 1810 and gives a complete history of the Senate up until LBJ is elected into the body.

LESSWRONG
LW