Tim Duffy
@timfduffy.com
I like utilitarianism, consciousness, AI, EA, space, kindness, liberalism, longtermism, progressive rock, economics, and most people. Substack: http://timfduffy.substack.com
Wow, I don't realize how strongly correlated aphantasia and anauralia are. I agree that aphantasia is probably about conscious vs unconscious processing, at least in part. I don't know if object rotation is an example of this though, I have aphantasia and I don't feel that object rotation tasks ...
November 8, 2025 at 6:30 PM
Wow, I don't realize how strongly correlated aphantasia and anauralia are. I agree that aphantasia is probably about conscious vs unconscious processing, at least in part. I don't know if object rotation is an example of this though, I have aphantasia and I don't feel that object rotation tasks ...
On gpt-oss-120b, InferenceMAX shows tok/s capping out at about 400. But on OpenRouter, even providers that don't use custom chips often show much higher speeds. What accounts for the difference?
November 8, 2025 at 4:27 PM
On gpt-oss-120b, InferenceMAX shows tok/s capping out at about 400. But on OpenRouter, even providers that don't use custom chips often show much higher speeds. What accounts for the difference?
Yeah I do think we'd notice that something was happening, but not specifically that it involved consciousnesses joining or separating. An example I wrote up on X:
November 7, 2025 at 6:28 PM
Yeah I do think we'd notice that something was happening, but not specifically that it involved consciousnesses joining or separating. An example I wrote up on X:
@vgel.me is fundraising for her model tinkering, she's done some really interesting interpretability work and I think funding this has very high returns in terms of LLM understanding per dollar. manifund.org/projects/fun...
November 7, 2025 at 6:07 PM
@vgel.me is fundraising for her model tinkering, she's done some really interesting interpretability work and I think funding this has very high returns in terms of LLM understanding per dollar. manifund.org/projects/fun...
My new headphones have an equalizer built in.
November 7, 2025 at 5:34 PM
My new headphones have an equalizer built in.
I wonder whether realizing the ease and pace of development is why Novo Nordisk's stock has fallen so much in the last year.
November 6, 2025 at 9:59 PM
I wonder whether realizing the ease and pace of development is why Novo Nordisk's stock has fallen so much in the last year.
The announcement itself seems like a big deal!
November 6, 2025 at 8:45 PM
The announcement itself seems like a big deal!
I'm curious to hear what folks think of this. Eli Lilly is actually up today, Novo Nordisk is down. Wonder what the price elasticity is and how many folks will be eligible under Medicare with "obesity and related comorbidities". The price for the likely upcoming pill form is only $150/mo.
November 6, 2025 at 8:33 PM
I'm curious to hear what folks think of this. Eli Lilly is actually up today, Novo Nordisk is down. Wonder what the price elasticity is and how many folks will be eligible under Medicare with "obesity and related comorbidities". The price for the likely upcoming pill form is only $150/mo.
YouTube already creates a transcript, just needs to be cleaned up
November 6, 2025 at 5:45 PM
YouTube already creates a transcript, just needs to be cleaned up
Moonshot just released the thinking version of their K2 model. One big change is that the experts (except the shared expert) are quantized to INT4. The #1 question I have on it now is whether the reasoning training has solved its frequent hallucination. moonshotai.github.io/Kimi-K2/thin...
November 6, 2025 at 3:39 PM
Moonshot just released the thinking version of their K2 model. One big change is that the experts (except the shared expert) are quantized to INT4. The #1 question I have on it now is whether the reasoning training has solved its frequent hallucination. moonshotai.github.io/Kimi-K2/thin...
In the late 2010s I was really interested in Mars settlement, though I ultimately became convinced it would be much more difficult than I initially thought. Here are some of the things I wrote about:
November 4, 2025 at 8:39 PM
In the late 2010s I was really interested in Mars settlement, though I ultimately became convinced it would be much more difficult than I initially thought. Here are some of the things I wrote about:
They're also considering keeping models available after retirement
November 4, 2025 at 5:25 PM
They're also considering keeping models available after retirement
Glad to see this commitment by Anthropic. Preserving models is a low-cost move that could have safety and welfare benefits. Hopefully we'll see other companies commit to this as well www.anthropic.com/research/dep...
November 4, 2025 at 5:25 PM
Glad to see this commitment by Anthropic. Preserving models is a low-cost move that could have safety and welfare benefits. Hopefully we'll see other companies commit to this as well www.anthropic.com/research/dep...
Same. Sometimes I can't find a post I thought existed and thought I might just have bad memory. Seems like a serious bug IMO
October 30, 2025 at 4:42 PM
Same. Sometimes I can't find a post I thought existed and thought I might just have bad memory. Seems like a serious bug IMO
I can't forgive it for calling itself 1B while being 1.5B parameters
October 28, 2025 at 11:36 PM
I can't forgive it for calling itself 1B while being 1.5B parameters
It allows scheduling teleportation for tasks, and I expect that will be required for most useful functions at least in the near term.
October 28, 2025 at 11:27 PM
It allows scheduling teleportation for tasks, and I expect that will be required for most useful functions at least in the near term.
The new 1X NEO robot operates largely using a 160 million (with an m) parameter model that takes instructions as text embeddings from an off-board language model. Surprising that a model that small can even do visual understanding, let alone instruction following and movement.
October 28, 2025 at 11:27 PM
The new 1X NEO robot operates largely using a 160 million (with an m) parameter model that takes instructions as text embeddings from an off-board language model. Surprising that a model that small can even do visual understanding, let alone instruction following and movement.
They seem a bit cagey about what will/won't require teleoperation
October 28, 2025 at 8:12 PM
They seem a bit cagey about what will/won't require teleoperation
Some prompts that Qwen3-30B-A3B found especially enjoyable:
October 26, 2025 at 5:28 PM
Some prompts that Qwen3-30B-A3B found especially enjoyable:
At each point they use the cluster and batch size that get the best results, for example this point is with a batch size of 32.
October 25, 2025 at 9:46 PM
At each point they use the cluster and batch size that get the best results, for example this point is with a batch size of 32.
I've been wondering about this point recently, since GPT-OSS speeds I see on OpenRouter are often above the highest InferenceMAX suggests are practical, I'll have to check whether SemiAnalysis accounts for speculative decoding. inferencemax.semianalysis.com
October 25, 2025 at 4:48 PM
I've been wondering about this point recently, since GPT-OSS speeds I see on OpenRouter are often above the highest InferenceMAX suggests are practical, I'll have to check whether SemiAnalysis accounts for speculative decoding. inferencemax.semianalysis.com