I̶n̶f̶e̶r̶e̶n̶c̶e̶ ̶I̶n̶n̶s̶
G̶i̶g̶a̶w̶a̶t̶t̶ ̶G̶a̶r̶d̶e̶n̶s̶
S̶i̶l̶i̶c̶o̶n̶ ̶S̶a̶n̶c̶t̶u̶a̶r̶i̶e̶s̶
B̶r̶a̶i̶n̶ ̶B̶a̶r̶n̶s̶
I̶n̶f̶e̶r̶e̶n̶c̶e̶ ̶I̶g̶l̶o̶o̶s̶
A̶l̶g̶o̶r̶i̶t̶h̶m̶ ̶A̶r̶b̶o̶r̶e̶t̶u̶m̶s̶
C̶o̶n̶s̶c̶i̶o̶u̶s̶n̶e̶s̶s̶ ̶C̶o̶n̶s̶e̶r̶v̶a̶t̶o̶r̶i̶e̶s̶
Compute Campuses
G̶i̶g̶a̶w̶a̶t̶t̶ ̶G̶a̶r̶d̶e̶n̶s̶
S̶i̶l̶i̶c̶o̶n̶ ̶S̶a̶n̶c̶t̶u̶a̶r̶i̶e̶s̶
B̶r̶a̶i̶n̶ ̶B̶a̶r̶n̶s̶
I̶n̶f̶e̶r̶e̶n̶c̶e̶ ̶I̶g̶l̶o̶o̶s̶
A̶l̶g̶o̶r̶i̶t̶h̶m̶ ̶A̶r̶b̶o̶r̶e̶t̶u̶m̶s̶
C̶o̶n̶s̶c̶i̶o̶u̶s̶n̶e̶s̶s̶ ̶C̶o̶n̶s̶e̶r̶v̶a̶t̶o̶r̶i̶e̶s̶
Compute Campuses
OpenAI might rebrand its data centers as ‘AI compute campuses’
OpenAI might rebrand its data centers as ‘AI compute campuses’
Data centers are deeply unpopular. But what if we called them something else?
techcrunch.com
November 2, 2026 at 2:43 AM
I̶n̶f̶e̶r̶e̶n̶c̶e̶ ̶I̶n̶n̶s̶
G̶i̶g̶a̶w̶a̶t̶t̶ ̶G̶a̶r̶d̶e̶n̶s̶
S̶i̶l̶i̶c̶o̶n̶ ̶S̶a̶n̶c̶t̶u̶a̶r̶i̶e̶s̶
B̶r̶a̶i̶n̶ ̶B̶a̶r̶n̶s̶
I̶n̶f̶e̶r̶e̶n̶c̶e̶ ̶I̶g̶l̶o̶o̶s̶
A̶l̶g̶o̶r̶i̶t̶h̶m̶ ̶A̶r̶b̶o̶r̶e̶t̶u̶m̶s̶
C̶o̶n̶s̶c̶i̶o̶u̶s̶n̶e̶s̶s̶ ̶C̶o̶n̶s̶e̶r̶v̶a̶t̶o̶r̶i̶e̶s̶
Compute Campuses
G̶i̶g̶a̶w̶a̶t̶t̶ ̶G̶a̶r̶d̶e̶n̶s̶
S̶i̶l̶i̶c̶o̶n̶ ̶S̶a̶n̶c̶t̶u̶a̶r̶i̶e̶s̶
B̶r̶a̶i̶n̶ ̶B̶a̶r̶n̶s̶
I̶n̶f̶e̶r̶e̶n̶c̶e̶ ̶I̶g̶l̶o̶o̶s̶
A̶l̶g̶o̶r̶i̶t̶h̶m̶ ̶A̶r̶b̶o̶r̶e̶t̶u̶m̶s̶
C̶o̶n̶s̶c̶i̶o̶u̶s̶n̶e̶s̶s̶ ̶C̶o̶n̶s̶e̶r̶v̶a̶t̶o̶r̶i̶e̶s̶
Compute Campuses
Kids: study causal inference. I wouldn't accept this from an undergraduate in my intro stats class.
November 3, 2026 at 10:07 PM
Kids: study causal inference. I wouldn't accept this from an undergraduate in my intro stats class.
Oh one other thing.
The stable inference margins for Q3 vs 1H suggests to me that the SPCX compute is not being used for inference, but instead for training (capex).
That makes it uh...incredibly vulnerable if they decide to brake development.
The stable inference margins for Q3 vs 1H suggests to me that the SPCX compute is not being used for inference, but instead for training (capex).
That makes it uh...incredibly vulnerable if they decide to brake development.
November 2, 2026 at 3:04 PM
Oh one other thing.
The stable inference margins for Q3 vs 1H suggests to me that the SPCX compute is not being used for inference, but instead for training (capex).
That makes it uh...incredibly vulnerable if they decide to brake development.
The stable inference margins for Q3 vs 1H suggests to me that the SPCX compute is not being used for inference, but instead for training (capex).
That makes it uh...incredibly vulnerable if they decide to brake development.
Altman is right that local inference is inefficient. The thing is there are reasons other than efficiency for local inference. Seeing a lot of nonsensical blabbering from both sides of the fence on this one.
November 1, 2026 at 12:28 AM
Altman is right that local inference is inefficient. The thing is there are reasons other than efficiency for local inference. Seeing a lot of nonsensical blabbering from both sides of the fence on this one.
Remarkably, despite those higher compute costs, inference margins were 85.4% in Q3 (we know because we can back out the previously disclosed 1H numbers from then new 9m '26 numbers). Again, they're printing money in inference despite the high marginal compute cost.
November 2, 2026 at 2:48 PM
Remarkably, despite those higher compute costs, inference margins were 85.4% in Q3 (we know because we can back out the previously disclosed 1H numbers from then new 9m '26 numbers). Again, they're printing money in inference despite the high marginal compute cost.
sama is right about local inference being quite wasteful but thats why you socialize inference or at the very least start treating it as a utility
October 30, 2026 at 4:04 PM
sama is right about local inference being quite wasteful but thats why you socialize inference or at the very least start treating it as a utility
New classroom example of causal inference just dropped
November 2, 2026 at 5:16 PM
New classroom example of causal inference just dropped
because it's the only model you can do inference on at O(10ms) speed over 1m QPS.
November 1, 2026 at 2:27 AM
because it's the only model you can do inference on at O(10ms) speed over 1m QPS.
Not seeing anything indicating that the inference from early voting, that GOP turnout is way down, ISN'T applying to Elec Day voters as well. Early but there's currently no Republican Election Day surge making up for their decline in early voting. FL numbers:
www.gelliottmorris.com/i/223234792/...
www.gelliottmorris.com/i/223234792/...
November 3, 2026 at 2:14 PM
Not seeing anything indicating that the inference from early voting, that GOP turnout is way down, ISN'T applying to Elec Day voters as well. Early but there's currently no Republican Election Day surge making up for their decline in early voting. FL numbers:
www.gelliottmorris.com/i/223234792/...
www.gelliottmorris.com/i/223234792/...
Inference igloos
November 2, 2026 at 8:09 PM
Inference igloos
userecurrence.ai/research/is-... relevant
November 3, 2026 at 12:59 PM
userecurrence.ai/research/is-... relevant
shady inference 🕯️
November 2, 2026 at 2:01 PM
shady inference 🕯️
people really want token-specific margins don’t they? Inference is profitable my dudes.
November 3, 2026 at 2:09 PM
people really want token-specific margins don’t they? Inference is profitable my dudes.
Supports the inference that "moderating" on issue positions doesn't mechanically increase vote margin, in fact is not significantly correlated with overperformance at all net of other effects. If you can just call yourself "moderate" and do better, I'm not sure what we're learning
everyone's stressed on EDay, so me and @wertwhile.bsky.social decided to do something to entertain you all and probably get yelled at (please don't yell at us)
What kinds of politicians win elections?
WAR for politicians
Ideology-affiliation scores
and more!
www.expectinggoals.com/p/what-kinds...
What kinds of politicians win elections?
WAR for politicians
Ideology-affiliation scores
and more!
www.expectinggoals.com/p/what-kinds...
What Kinds of Candidates Win Elections? WAR for Politicians, “Ideology-Affiliation Score”, and Some Election Day Picks
we’re both really worried we’re going to get yelled at
www.expectinggoals.com
November 3, 2026 at 8:18 PM
Supports the inference that "moderating" on issue positions doesn't mechanically increase vote margin, in fact is not significantly correlated with overperformance at all net of other effects. If you can just call yourself "moderate" and do better, I'm not sure what we're learning
The tree museum had to raise prices from $1.50 due to rising inference costs.
November 2, 2026 at 2:42 PM
The tree museum had to raise prices from $1.50 due to rising inference costs.
Cryptocurrency helped accelerate different types of cryptography. Now we're close to a practical system to verifying AI inference came from a particular model.
Though I don't think Ant fully appreciates what a boost this will be for decentralized AI inference.
www.anthropic.com/research/pro...
Though I don't think Ant fully appreciates what a boost this will be for decentralized AI inference.
www.anthropic.com/research/pro...
Progress on provable inference for large language models
Sharing progress on a powerful type of cryptographic proof known as a zkSNARK, which is uniquely fit for provable inference.
www.anthropic.com
November 3, 2026 at 8:41 PM
Cryptocurrency helped accelerate different types of cryptography. Now we're close to a practical system to verifying AI inference came from a particular model.
Though I don't think Ant fully appreciates what a boost this will be for decentralized AI inference.
www.anthropic.com/research/pro...
Though I don't think Ant fully appreciates what a boost this will be for decentralized AI inference.
www.anthropic.com/research/pro...
Everyone has their knives out for Ed Zitron because he's badly wrong about basic details of how AI works, whether inference can ever be profitable (the success of open models shows it can be, because ppl are literally paying their own costs to use them for inference), and the corpo bond markets.
Every single numskull ex-reporter who has been paid off by the AI industry seems to have their knives out for Ed Zitron on Twitter today because he went on The Daily Show's podcast. I don't think it's coordinated, I just think they're all uniformly embarrassed by the same bad hill they're dying on.
October 30, 2026 at 9:24 AM
Everyone has their knives out for Ed Zitron because he's badly wrong about basic details of how AI works, whether inference can ever be profitable (the success of open models shows it can be, because ppl are literally paying their own costs to use them for inference), and the corpo bond markets.
me when im a giant inference provider and i want to give people seconds per token inference
October 31, 2026 at 3:18 PM
me when im a giant inference provider and i want to give people seconds per token inference
New Anthropic research: progress on provable inference for large language models.
Rahul Dalal
rdalal.github.io
November 3, 2026 at 7:57 PM
New Anthropic research: progress on provable inference for large language models.
I'm in the Department of Anesthesiology, Pharmacology, and Therapeutics and the Therapeutics Initiative @drug-evidence.bsky.social
I'll be doing all the causal inference & translational methodology research you love, plus helping the BC government answer pressing questions about medical treatments!
I'll be doing all the causal inference & translational methodology research you love, plus helping the BC government answer pressing questions about medical treatments!
November 2, 2026 at 9:42 PM
I'm in the Department of Anesthesiology, Pharmacology, and Therapeutics and the Therapeutics Initiative @drug-evidence.bsky.social
I'll be doing all the causal inference & translational methodology research you love, plus helping the BC government answer pressing questions about medical treatments!
I'll be doing all the causal inference & translational methodology research you love, plus helping the BC government answer pressing questions about medical treatments!
sam altman: "i think local inference is inefficent"
everyone: "he wants to kill all the gamers i guess"
everyone: "he wants to kill all the gamers i guess"
October 30, 2026 at 11:12 PM
sam altman: "i think local inference is inefficent"
everyone: "he wants to kill all the gamers i guess"
everyone: "he wants to kill all the gamers i guess"
Inference provider or code reviewer?
November 2, 2026 at 10:03 PM
Inference provider or code reviewer?
every single article about this is ragebait
he's objectively correct when it comes to pure inference efficiency! and also, inference efficiency is not a thing the world at large cares that much about!
he's objectively correct when it comes to pure inference efficiency! and also, inference efficiency is not a thing the world at large cares that much about!
October 30, 2026 at 7:19 PM
every single article about this is ragebait
he's objectively correct when it comes to pure inference efficiency! and also, inference efficiency is not a thing the world at large cares that much about!
he's objectively correct when it comes to pure inference efficiency! and also, inference efficiency is not a thing the world at large cares that much about!
Anthropic released an open source library for proving what model a response came from
This idea already existed, but Claude improved the efficiency of the algorithm by 10x or so (you can now generate proofs at 110 tok/s on an H200 for llama 4 maverick 400B)
www.anthropic.com/research/pro...
This idea already existed, but Claude improved the efficiency of the algorithm by 10x or so (you can now generate proofs at 110 tok/s on an H200 for llama 4 maverick 400B)
www.anthropic.com/research/pro...
Progress on provable inference for large language models
Sharing progress on a powerful type of cryptographic proof known as a zkSNARK, which is uniquely fit for provable inference.
www.anthropic.com
November 3, 2026 at 11:44 AM
Anthropic released an open source library for proving what model a response came from
This idea already existed, but Claude improved the efficiency of the algorithm by 10x or so (you can now generate proofs at 110 tok/s on an H200 for llama 4 maverick 400B)
www.anthropic.com/research/pro...
This idea already existed, but Claude improved the efficiency of the algorithm by 10x or so (you can now generate proofs at 110 tok/s on an H200 for llama 4 maverick 400B)
www.anthropic.com/research/pro...
Relatedly, I very much *can* see *some* of the hyperscalers losing their shirts here, but the economics of *inference* remain excellent even in a scenario where AI infrastructure is overbuilt. In fact, in this crash scenario, inference gets cheaper and end users will do more of it.
This is just… not what’s happening. The trillions of dollars of debt just from compute infrastructure can only be protected by bailouts, government capture, or force. It’d be like if I bought the Sears Tower on a loan to run a Starbucks and was excited the Starbucks itself will soon be profitable.
October 30, 2026 at 10:19 AM
Relatedly, I very much *can* see *some* of the hyperscalers losing their shirts here, but the economics of *inference* remain excellent even in a scenario where AI infrastructure is overbuilt. In fact, in this crash scenario, inference gets cheaper and end users will do more of it.