do me a favor and set a calendar event or reminder for this post one year from now
September 15, 2026 at 5:11 PM
do me a favor and set a calendar event or reminder for this post one year from now
i think you underestimate code, and underestimate the power of cheap code.
also take a look at what’s happening in math, science, healthcare, chip design.
update your priors, yours are assuredly out of date
also take a look at what’s happening in math, science, healthcare, chip design.
update your priors, yours are assuredly out of date
September 15, 2026 at 5:10 PM
i think you underestimate code, and underestimate the power of cheap code.
also take a look at what’s happening in math, science, healthcare, chip design.
update your priors, yours are assuredly out of date
also take a look at what’s happening in math, science, healthcare, chip design.
update your priors, yours are assuredly out of date
i’ve been a software engineer for 22 years. today about 1% of my code is hand written. a year ago that was unthinkable.
we’re communicating via AI generated code. fyi.
we’re communicating via AI generated code. fyi.
September 15, 2026 at 4:59 PM
i’ve been a software engineer for 22 years. today about 1% of my code is hand written. a year ago that was unthinkable.
we’re communicating via AI generated code. fyi.
we’re communicating via AI generated code. fyi.
“under-delivered” is an insane take considering that the latest models are saturating benchmarks and have completely changed entire professions much faster than anyone expected.
September 15, 2026 at 4:52 PM
“under-delivered” is an insane take considering that the latest models are saturating benchmarks and have completely changed entire professions much faster than anyone expected.
prove some math. it’s surprisingly fun
September 12, 2026 at 1:57 PM
prove some math. it’s surprisingly fun
i’m switching my entire company from cursor to codex. the plans are too damned generous and openai models are too token efficient.
astra, in particular, is excellent at lower reasoning levels.
using cost per task it’s financially irresponsible to use anthropic
astra, in particular, is excellent at lower reasoning levels.
using cost per task it’s financially irresponsible to use anthropic
September 11, 2026 at 5:11 PM
i’m switching my entire company from cursor to codex. the plans are too damned generous and openai models are too token efficient.
astra, in particular, is excellent at lower reasoning levels.
using cost per task it’s financially irresponsible to use anthropic
astra, in particular, is excellent at lower reasoning levels.
using cost per task it’s financially irresponsible to use anthropic
I’ve been thinking a lot lately about what it would take for an LLM to be a competent drafter. Probably explicit RL, especially on novel card evaluation.
Still, this is farther than I expected we’d get in 2026.
Still, this is farther than I expected we’d get in 2026.
September 8, 2026 at 1:06 PM
I’ve been thinking a lot lately about what it would take for an LLM to be a competent drafter. Probably explicit RL, especially on novel card evaluation.
Still, this is farther than I expected we’d get in 2026.
Still, this is farther than I expected we’d get in 2026.
*accidentally produce
you could create a substantial copy but you’d have to try so hard the LLM would also copy the license
you could create a substantial copy but you’d have to try so hard the LLM would also copy the license
September 7, 2026 at 4:15 PM
*accidentally produce
you could create a substantial copy but you’d have to try so hard the LLM would also copy the license
you could create a substantial copy but you’d have to try so hard the LLM would also copy the license
that only applies for substantial copies. substantial copies are basically impossible to produce with an LLM.
i would recommend reading the MIT license
i would recommend reading the MIT license
September 7, 2026 at 4:13 PM
that only applies for substantial copies. substantial copies are basically impossible to produce with an LLM.
i would recommend reading the MIT license
i would recommend reading the MIT license
have you ever read an open software license? or go to wikipedia and look up "MIT license" or "apache license"
September 7, 2026 at 3:25 PM
have you ever read an open software license? or go to wikipedia and look up "MIT license" or "apache license"
i have probably tens of thousands of lines of code in LLM training corpuses and i'm actively glad of that fact.
i'd rather LLMs code like me than anyone else, so yes please ingest my hard work. i MIT license it for a reason.
i'd rather LLMs code like me than anyone else, so yes please ingest my hard work. i MIT license it for a reason.
September 7, 2026 at 3:25 PM
i have probably tens of thousands of lines of code in LLM training corpuses and i'm actively glad of that fact.
i'd rather LLMs code like me than anyone else, so yes please ingest my hard work. i MIT license it for a reason.
i'd rather LLMs code like me than anyone else, so yes please ingest my hard work. i MIT license it for a reason.
hey just jumping in here to provide independent verification that you don't know what you're talking about.
before engaging in further commentary on the internet please learn what the hell you're talking about.
and preferably just shut up permanently, if that's where you're going with this.
before engaging in further commentary on the internet please learn what the hell you're talking about.
and preferably just shut up permanently, if that's where you're going with this.
September 6, 2026 at 12:34 PM
hey just jumping in here to provide independent verification that you don't know what you're talking about.
before engaging in further commentary on the internet please learn what the hell you're talking about.
and preferably just shut up permanently, if that's where you're going with this.
before engaging in further commentary on the internet please learn what the hell you're talking about.
and preferably just shut up permanently, if that's where you're going with this.
"highly inappropriate tools in contexts where truth and correctness are of utmost importance"
wait til that dude finds out about Lean Pool or Tau Ceti or the FLT proof, etc
wait til that dude finds out about Lean Pool or Tau Ceti or the FLT proof, etc
September 6, 2026 at 12:31 PM
"highly inappropriate tools in contexts where truth and correctness are of utmost importance"
wait til that dude finds out about Lean Pool or Tau Ceti or the FLT proof, etc
wait til that dude finds out about Lean Pool or Tau Ceti or the FLT proof, etc
a lot of ad hominem in there for sure.
it’s absurd that they didn’t have network monitoring and more restrictive egress on the artifactory server.
but that’s about all it is, yeah? garden variety incompetence while wielding a dangerous system
it’s absurd that they didn’t have network monitoring and more restrictive egress on the artifactory server.
but that’s about all it is, yeah? garden variety incompetence while wielding a dangerous system
August 30, 2026 at 7:51 PM
a lot of ad hominem in there for sure.
it’s absurd that they didn’t have network monitoring and more restrictive egress on the artifactory server.
but that’s about all it is, yeah? garden variety incompetence while wielding a dangerous system
it’s absurd that they didn’t have network monitoring and more restrictive egress on the artifactory server.
but that’s about all it is, yeah? garden variety incompetence while wielding a dangerous system
there’s a reason why microsoft azure is huge and it’s not because it’s good or cheap
these same market forces will transfer to AI spend
these same market forces will transfer to AI spend
August 22, 2026 at 12:43 PM
there’s a reason why microsoft azure is huge and it’s not because it’s good or cheap
these same market forces will transfer to AI spend
these same market forces will transfer to AI spend
enterprises, and most companies broadly, will not want to manage their own hardware.
especially if you have regulatory concerns (e.g. hipaa) you’re going to move towards providers that can offer you assurances (e.g. a BAA)
especially if you have regulatory concerns (e.g. hipaa) you’re going to move towards providers that can offer you assurances (e.g. a BAA)
August 22, 2026 at 12:41 PM
enterprises, and most companies broadly, will not want to manage their own hardware.
especially if you have regulatory concerns (e.g. hipaa) you’re going to move towards providers that can offer you assurances (e.g. a BAA)
especially if you have regulatory concerns (e.g. hipaa) you’re going to move towards providers that can offer you assurances (e.g. a BAA)
i use Sol for a lot but not for communication. its strength is in dense communication intended for itself. i honestly wonder if an LLM is capable of both dense thinking and good communication.
fable is the best model at communicating but it’s still not “good,” which is the real problem
fable is the best model at communicating but it’s still not “good,” which is the real problem
August 16, 2026 at 7:09 PM
i use Sol for a lot but not for communication. its strength is in dense communication intended for itself. i honestly wonder if an LLM is capable of both dense thinking and good communication.
fable is the best model at communicating but it’s still not “good,” which is the real problem
fable is the best model at communicating but it’s still not “good,” which is the real problem
i would recommend you start reading the prompts the orchestrator is feeding subagents. i have been routinely horrified and abandoned several plans. this happens less often with smarter models.
August 15, 2026 at 4:53 PM
i would recommend you start reading the prompts the orchestrator is feeding subagents. i have been routinely horrified and abandoned several plans. this happens less often with smarter models.
i don’t want to get into too many specifics that aren’t generalizable but: my experience is quite different.
maybe it’s a difference of harness (i use cursor) or in model preference (gpt-5.6 family for most things)
maybe it’s a difference of harness (i use cursor) or in model preference (gpt-5.6 family for most things)
August 15, 2026 at 4:48 PM
i don’t want to get into too many specifics that aren’t generalizable but: my experience is quite different.
maybe it’s a difference of harness (i use cursor) or in model preference (gpt-5.6 family for most things)
maybe it’s a difference of harness (i use cursor) or in model preference (gpt-5.6 family for most things)
the way i basically think about it is if the work can be validated by a more expensive model, use the cheaper one.
the only reason i use a cheap model for implementation is because an expensive one does adversarial review
i also have a good budget at work which is certainly a factor
the only reason i use a cheap model for implementation is because an expensive one does adversarial review
i also have a good budget at work which is certainly a factor
August 15, 2026 at 4:45 PM
the way i basically think about it is if the work can be validated by a more expensive model, use the cheaper one.
the only reason i use a cheap model for implementation is because an expensive one does adversarial review
i also have a good budget at work which is certainly a factor
the only reason i use a cheap model for implementation is because an expensive one does adversarial review
i also have a good budget at work which is certainly a factor
Turning plans into agent instructions and validating their work is something that smart models are meaningfully better at IMO
idk how much you’re saving yourself by micro-slicing roles either
idk how much you’re saving yourself by micro-slicing roles either
August 15, 2026 at 4:36 PM
Turning plans into agent instructions and validating their work is something that smart models are meaningfully better at IMO
idk how much you’re saving yourself by micro-slicing roles either
idk how much you’re saving yourself by micro-slicing roles either
Ehh, idk. I find having a smart model gather context about the solution to be rather important.
Dumb models are way more likely to propose a narrow solution. Planning and orchestrating are where thinking broadly is helpful.
Dumb models are way more likely to propose a narrow solution. Planning and orchestrating are where thinking broadly is helpful.
August 15, 2026 at 4:33 PM
Ehh, idk. I find having a smart model gather context about the solution to be rather important.
Dumb models are way more likely to propose a narrow solution. Planning and orchestrating are where thinking broadly is helpful.
Dumb models are way more likely to propose a narrow solution. Planning and orchestrating are where thinking broadly is helpful.