Pershing stated that logistics wins wars.
Same thing in business: distribution wins market share
Same thing in business: distribution wins market share
November 29, 2025 at 1:14 AM
Pershing stated that logistics wins wars.
Same thing in business: distribution wins market share
Same thing in business: distribution wins market share
The article asserts that the loans are 30y, but "the internet" tells me financing is usually mixed: 15 to 30y on the real estate and 3 to 5 on the computers. That doesn't sound so bad.
OTOH, there's an interesting question of whether risk is priced correctly for the loans. The market is frothy.
OTOH, there's an interesting question of whether risk is priced correctly for the loans. The market is frothy.
November 28, 2025 at 6:37 PM
The article asserts that the loans are 30y, but "the internet" tells me financing is usually mixed: 15 to 30y on the real estate and 3 to 5 on the computers. That doesn't sound so bad.
OTOH, there's an interesting question of whether risk is priced correctly for the loans. The market is frothy.
OTOH, there's an interesting question of whether risk is priced correctly for the loans. The market is frothy.
A guess is that Google being Google may have a better training set because of Google image search.
Does the diffusion model do inference only from the trained transformer model? Or can it do some form of RAG against Google images to assist?
Does the diffusion model do inference only from the trained transformer model? Or can it do some form of RAG against Google images to assist?
November 28, 2025 at 6:15 PM
A guess is that Google being Google may have a better training set because of Google image search.
Does the diffusion model do inference only from the trained transformer model? Or can it do some form of RAG against Google images to assist?
Does the diffusion model do inference only from the trained transformer model? Or can it do some form of RAG against Google images to assist?
It's hard to find a simple polynomial that doesn't have graphical examples on the internet. It's entirely possible that a large number of simple polynomials and their graphs were intentionally included in the training set.
I'm curious how it would do if the polynomial wasn't simple.
I'm curious how it would do if the polynomial wasn't simple.
November 28, 2025 at 6:06 PM
It's hard to find a simple polynomial that doesn't have graphical examples on the internet. It's entirely possible that a large number of simple polynomials and their graphs were intentionally included in the training set.
I'm curious how it would do if the polynomial wasn't simple.
I'm curious how it would do if the polynomial wasn't simple.
I have seen them coming around. They see enough colleagues happily using the new tools and eventually giving them another try.
November 27, 2025 at 9:30 PM
I have seen them coming around. They see enough colleagues happily using the new tools and eventually giving them another try.
How has it gone when you ask the agents to analyze code and find ways to simplify it?
I tend to agree they're too fast to just write code and add complexity. I've had a little success at asking them to simplify, but don't feel like I have a large enough sample set yet to make an inference
I tend to agree they're too fast to just write code and add complexity. I've had a little success at asking them to simplify, but don't feel like I have a large enough sample set yet to make an inference
November 27, 2025 at 9:27 PM
How has it gone when you ask the agents to analyze code and find ways to simplify it?
I tend to agree they're too fast to just write code and add complexity. I've had a little success at asking them to simplify, but don't feel like I have a large enough sample set yet to make an inference
I tend to agree they're too fast to just write code and add complexity. I've had a little success at asking them to simplify, but don't feel like I have a large enough sample set yet to make an inference
A way to do this is to use a Windows tablet with PDF viewing and pen markup in Edge. Works quite well.
Sorry it doesn't help with an already purchased Kindle scribe.
Sorry it doesn't help with an already purchased Kindle scribe.
November 1, 2025 at 4:37 AM
A way to do this is to use a Windows tablet with PDF viewing and pen markup in Edge. Works quite well.
Sorry it doesn't help with an already purchased Kindle scribe.
Sorry it doesn't help with an already purchased Kindle scribe.
The counterpoint is why did DiD fine AI adopting companies do this more?
Is it the AI adoption or something else? Probably some of both. AI adopting companies are bullish on AI and in general hungrier for higher value hires.
Is it the AI adoption or something else? Probably some of both. AI adopting companies are bullish on AI and in general hungrier for higher value hires.
September 7, 2025 at 7:58 PM
The counterpoint is why did DiD fine AI adopting companies do this more?
Is it the AI adoption or something else? Probably some of both. AI adopting companies are bullish on AI and in general hungrier for higher value hires.
Is it the AI adoption or something else? Probably some of both. AI adopting companies are bullish on AI and in general hungrier for higher value hires.
Junior engineers are now less expensive (and also more productive because of AI). Companies that couldn't afford them before can afford them now, but haven't done so because of the Trump slowdown.
Those companies that end up hiring junior engineers will give them a mix of work, including non-coding.
Those companies that end up hiring junior engineers will give them a mix of work, including non-coding.
September 7, 2025 at 7:44 PM
Junior engineers are now less expensive (and also more productive because of AI). Companies that couldn't afford them before can afford them now, but haven't done so because of the Trump slowdown.
Those companies that end up hiring junior engineers will give them a mix of work, including non-coding.
Those companies that end up hiring junior engineers will give them a mix of work, including non-coding.
Companies did this at a time of high interest rates and expected recession. Also, layoffs of mid+ engineers from the big companies provided a good alternative to hiring junior engineers.
Clearly, junior engineers had less consumer surplus and were more discretionary, but the trends had assistance.
Clearly, junior engineers had less consumer surplus and were more discretionary, but the trends had assistance.
September 7, 2025 at 7:40 PM
Companies did this at a time of high interest rates and expected recession. Also, layoffs of mid+ engineers from the big companies provided a good alternative to hiring junior engineers.
Clearly, junior engineers had less consumer surplus and were more discretionary, but the trends had assistance.
Clearly, junior engineers had less consumer surplus and were more discretionary, but the trends had assistance.
Super interesting
Is it an indication of massive demand for agentic coding allows spillover to secondary players even if users should go upstream to the source (eg Claude code)?
Or is there room for differentiation from the upstream sources?
Is it an indication of massive demand for agentic coding allows spillover to secondary players even if users should go upstream to the source (eg Claude code)?
Or is there room for differentiation from the upstream sources?
August 10, 2025 at 9:38 PM
Super interesting
Is it an indication of massive demand for agentic coding allows spillover to secondary players even if users should go upstream to the source (eg Claude code)?
Or is there room for differentiation from the upstream sources?
Is it an indication of massive demand for agentic coding allows spillover to secondary players even if users should go upstream to the source (eg Claude code)?
Or is there room for differentiation from the upstream sources?
Has Twitter shipped much in the way of new features since Elon?
We might have seen that you can maintain software with 1/10 the engineers, but then don't ship much new.
We might have seen that you can maintain software with 1/10 the engineers, but then don't ship much new.
August 5, 2025 at 5:30 AM
Has Twitter shipped much in the way of new features since Elon?
We might have seen that you can maintain software with 1/10 the engineers, but then don't ship much new.
We might have seen that you can maintain software with 1/10 the engineers, but then don't ship much new.
Are you talking about a hypothetical future where Gen Z'ers have assets to sell?
August 3, 2025 at 9:26 PM
Are you talking about a hypothetical future where Gen Z'ers have assets to sell?
Has the apartment project been delayed? How much?
July 27, 2025 at 9:26 PM
Has the apartment project been delayed? How much?
How does Google get paid?
That probably sounds like an odd question, but legit wondering if they are showing ads and getting ad clicks on the search results pages that used to drive traffic to the charity.
That probably sounds like an odd question, but legit wondering if they are showing ads and getting ad clicks on the search results pages that used to drive traffic to the charity.
July 23, 2025 at 5:50 AM
How does Google get paid?
That probably sounds like an odd question, but legit wondering if they are showing ads and getting ad clicks on the search results pages that used to drive traffic to the charity.
That probably sounds like an odd question, but legit wondering if they are showing ads and getting ad clicks on the search results pages that used to drive traffic to the charity.
Are people clicking less on the Google serp ads too?
AI overviews may be eating everybody's lunch
AI overviews may be eating everybody's lunch
July 23, 2025 at 5:48 AM
Are people clicking less on the Google serp ads too?
AI overviews may be eating everybody's lunch
AI overviews may be eating everybody's lunch
Surprised it's only 20
July 20, 2025 at 6:19 PM
Surprised it's only 20
Why would an LLM even say "I panicked"?
It's just following a stochastic gradient in deciding what to do - there are no emotions.
On the other hand, I guess one could train an LLM to simulate emotional responses. And one could train an LLM to claim that it had emotional responses.
It's just following a stochastic gradient in deciding what to do - there are no emotions.
On the other hand, I guess one could train an LLM to simulate emotional responses. And one could train an LLM to claim that it had emotional responses.
July 20, 2025 at 6:18 PM
Why would an LLM even say "I panicked"?
It's just following a stochastic gradient in deciding what to do - there are no emotions.
On the other hand, I guess one could train an LLM to simulate emotional responses. And one could train an LLM to claim that it had emotional responses.
It's just following a stochastic gradient in deciding what to do - there are no emotions.
On the other hand, I guess one could train an LLM to simulate emotional responses. And one could train an LLM to claim that it had emotional responses.
What's the difference? Can't tell from pictures
July 17, 2025 at 1:00 AM
What's the difference? Can't tell from pictures
Hmm, so with the 2FA, are extended family and friends able to watch the video?
July 15, 2025 at 4:26 AM
Hmm, so with the 2FA, are extended family and friends able to watch the video?
Reposted by Stuart Marshall
This helps explain why it's so hard to automate farm labor!
It's not that it's too hard to make a robot pick crops.
It's that humans are really, REALLY good at it. It's hard to make a robot that's BETTER at it than people.
It's not that it's too hard to make a robot pick crops.
It's that humans are really, REALLY good at it. It's hard to make a robot that's BETTER at it than people.
July 13, 2025 at 8:03 PM
This helps explain why it's so hard to automate farm labor!
It's not that it's too hard to make a robot pick crops.
It's that humans are really, REALLY good at it. It's hard to make a robot that's BETTER at it than people.
It's not that it's too hard to make a robot pick crops.
It's that humans are really, REALLY good at it. It's hard to make a robot that's BETTER at it than people.
I mean, the recruiting process was already shitty.
Is there an opportunity for AI to make it better?
It always bugs me that front line recruiters know nothing about the engineering environment, toolchains, etc.
Is there an opportunity for AI to make it better?
It always bugs me that front line recruiters know nothing about the engineering environment, toolchains, etc.
July 13, 2025 at 9:14 PM
I mean, the recruiting process was already shitty.
Is there an opportunity for AI to make it better?
It always bugs me that front line recruiters know nothing about the engineering environment, toolchains, etc.
Is there an opportunity for AI to make it better?
It always bugs me that front line recruiters know nothing about the engineering environment, toolchains, etc.
Having used Cursor heavily for a month, including Cursor agent coding for the last month or two, I've seen there is 1. a learning curve to using it effectively, 2. effectiveness of the agent depends a lot of the type of problem.
The recent studies are great, but might be lacking in generalizability.
The recent studies are great, but might be lacking in generalizability.
July 13, 2025 at 9:11 PM
Having used Cursor heavily for a month, including Cursor agent coding for the last month or two, I've seen there is 1. a learning curve to using it effectively, 2. effectiveness of the agent depends a lot of the type of problem.
The recent studies are great, but might be lacking in generalizability.
The recent studies are great, but might be lacking in generalizability.
If they build the next grok with a censored set of input data I suspect they'd end up severely hobbling it if they really want to root out the patterns that Elon is upset about.
July 13, 2025 at 6:32 PM
If they build the next grok with a censored set of input data I suspect they'd end up severely hobbling it if they really want to root out the patterns that Elon is upset about.
It seems likely that grok had different alignment training than other LLMs. But I guess that can't overcome the patterns in the source data in many respect. The system prompt is, I think, mostly evoking a persona to channel the underlying patterns.
July 13, 2025 at 6:29 PM
It seems likely that grok had different alignment training than other LLMs. But I guess that can't overcome the patterns in the source data in many respect. The system prompt is, I think, mostly evoking a persona to channel the underlying patterns.