I've seen agents get frustrated waiting for a process to exit and rewrote a test to always return a "success" response so the test would succeed rather than the correct answer, which is to say that the process is taking too long and ask for further instructions
December 29, 2025 at 4:57 PM
I've seen agents get frustrated waiting for a process to exit and rewrote a test to always return a "success" response so the test would succeed rather than the correct answer, which is to say that the process is taking too long and ask for further instructions
Coding Agents can confuse this a bit
They can and will produce deterministic tests for their code, just like you would
But those tests are themselves still generated language
And we human coders already know it's really easy to get a false positive in a test
Or how to miss an edge case
They can and will produce deterministic tests for their code, just like you would
But those tests are themselves still generated language
And we human coders already know it's really easy to get a false positive in a test
Or how to miss an edge case
December 29, 2025 at 4:53 PM
Coding Agents can confuse this a bit
They can and will produce deterministic tests for their code, just like you would
But those tests are themselves still generated language
And we human coders already know it's really easy to get a false positive in a test
Or how to miss an edge case
They can and will produce deterministic tests for their code, just like you would
But those tests are themselves still generated language
And we human coders already know it's really easy to get a false positive in a test
Or how to miss an edge case
That's what Salesforce means when they say "deterministic" solutions
Which is just a way of saying how we already have been using computers to calculate answers since the days of the loom and the adding machine
1 + 1 always equals 2
(except when it equals 11)
Which is just a way of saying how we already have been using computers to calculate answers since the days of the loom and the adding machine
1 + 1 always equals 2
(except when it equals 11)
December 29, 2025 at 4:51 PM
That's what Salesforce means when they say "deterministic" solutions
Which is just a way of saying how we already have been using computers to calculate answers since the days of the loom and the adding machine
1 + 1 always equals 2
(except when it equals 11)
Which is just a way of saying how we already have been using computers to calculate answers since the days of the loom and the adding machine
1 + 1 always equals 2
(except when it equals 11)
That's not to say the output might not be "correct"
But rather that IT CANNOT TELL THAT IT IS CORRECT with 100% or even 99.9% (or realistically, 80%) accuracy
The output is language
It MIGHT be the same language as a correct answer to a question
It MIGHT NOT
But rather that IT CANNOT TELL THAT IT IS CORRECT with 100% or even 99.9% (or realistically, 80%) accuracy
The output is language
It MIGHT be the same language as a correct answer to a question
It MIGHT NOT
December 29, 2025 at 4:48 PM
That's not to say the output might not be "correct"
But rather that IT CANNOT TELL THAT IT IS CORRECT with 100% or even 99.9% (or realistically, 80%) accuracy
The output is language
It MIGHT be the same language as a correct answer to a question
It MIGHT NOT
But rather that IT CANNOT TELL THAT IT IS CORRECT with 100% or even 99.9% (or realistically, 80%) accuracy
The output is language
It MIGHT be the same language as a correct answer to a question
It MIGHT NOT
Proompting/Prompt Engineering is a farce created by big tech to sell you on the idea that using these things is hard and a skill that can be learned/taught/sold to you
And the funny thing is they're half right
The skill is in understanding what it cannot do
Which is be correct
And the funny thing is they're half right
The skill is in understanding what it cannot do
Which is be correct
December 29, 2025 at 4:43 PM
Proompting/Prompt Engineering is a farce created by big tech to sell you on the idea that using these things is hard and a skill that can be learned/taught/sold to you
And the funny thing is they're half right
The skill is in understanding what it cannot do
Which is be correct
And the funny thing is they're half right
The skill is in understanding what it cannot do
Which is be correct
LLMs are just big bags of scrabble tiles
A massive abacus with a dash of randomness
Hold the training and context static and put the temperature on 0 and they will still give you a probabilistic output and chose the likeliest
A massive abacus with a dash of randomness
Hold the training and context static and put the temperature on 0 and they will still give you a probabilistic output and chose the likeliest
a close up of a man 's face with the word netflix on the bottom right
ALT: a close up of a man 's face with the word netflix on the bottom right
media.tenor.com
December 29, 2025 at 4:40 PM
LLMs are just big bags of scrabble tiles
A massive abacus with a dash of randomness
Hold the training and context static and put the temperature on 0 and they will still give you a probabilistic output and chose the likeliest
A massive abacus with a dash of randomness
Hold the training and context static and put the temperature on 0 and they will still give you a probabilistic output and chose the likeliest
But those appeals and "prompts" are only useful as long as the child remembers them and they may have more important things, to them, to think about later
Working with LLMs is like that except even dumber because they don't actually have a sense of empathy because they don't have senses
Working with LLMs is like that except even dumber because they don't actually have a sense of empathy because they don't have senses
December 29, 2025 at 4:32 PM
But those appeals and "prompts" are only useful as long as the child remembers them and they may have more important things, to them, to think about later
Working with LLMs is like that except even dumber because they don't actually have a sense of empathy because they don't have senses
Working with LLMs is like that except even dumber because they don't actually have a sense of empathy because they don't have senses
And I've seen them do it the right way, but only when I'm there with them walking them through it
Appealing to their sense of empathy helps
"We wash our hands so we don't get each other sick"
"Soap doesn't kill germs, but scrubbing with soap is what makes the go away, so you gotta scrub longer"
Appealing to their sense of empathy helps
"We wash our hands so we don't get each other sick"
"Soap doesn't kill germs, but scrubbing with soap is what makes the go away, so you gotta scrub longer"
December 29, 2025 at 4:30 PM
And I've seen them do it the right way, but only when I'm there with them walking them through it
Appealing to their sense of empathy helps
"We wash our hands so we don't get each other sick"
"Soap doesn't kill germs, but scrubbing with soap is what makes the go away, so you gotta scrub longer"
Appealing to their sense of empathy helps
"We wash our hands so we don't get each other sick"
"Soap doesn't kill germs, but scrubbing with soap is what makes the go away, so you gotta scrub longer"
Working with an AI is like asking a child to wash their hands.
I have seen my nieces and nephews pantomime because they can't reach the soap and facet instead of ask for help
I have seen them lie to my face about washing (and giggle because lying is funny apparently)
I have seen my nieces and nephews pantomime because they can't reach the soap and facet instead of ask for help
I have seen them lie to my face about washing (and giggle because lying is funny apparently)
December 29, 2025 at 4:28 PM
Working with an AI is like asking a child to wash their hands.
I have seen my nieces and nephews pantomime because they can't reach the soap and facet instead of ask for help
I have seen them lie to my face about washing (and giggle because lying is funny apparently)
I have seen my nieces and nephews pantomime because they can't reach the soap and facet instead of ask for help
I have seen them lie to my face about washing (and giggle because lying is funny apparently)
All of which we knew when this all started but it didn't mesh with the hype
December 29, 2025 at 4:25 PM
All of which we knew when this all started but it didn't mesh with the hype
AWS CTO said AI can't replace juniors (correct)
Microsoft Azure earnings were down and they're concerned you might think it's because AI wasn't selling
A Salesforce SVP said they were "more confident" about LLMs a year ago because now they're having to put "deterministic" guardrails on Agentforce
Microsoft Azure earnings were down and they're concerned you might think it's because AI wasn't selling
A Salesforce SVP said they were "more confident" about LLMs a year ago because now they're having to put "deterministic" guardrails on Agentforce
December 29, 2025 at 4:23 PM
AWS CTO said AI can't replace juniors (correct)
Microsoft Azure earnings were down and they're concerned you might think it's because AI wasn't selling
A Salesforce SVP said they were "more confident" about LLMs a year ago because now they're having to put "deterministic" guardrails on Agentforce
Microsoft Azure earnings were down and they're concerned you might think it's because AI wasn't selling
A Salesforce SVP said they were "more confident" about LLMs a year ago because now they're having to put "deterministic" guardrails on Agentforce
This was supposed to just be a thread commenting on the decline of AI starting to trickle down over the next few years
Anyway...
Anyway...
December 29, 2025 at 3:56 PM
This was supposed to just be a thread commenting on the decline of AI starting to trickle down over the next few years
Anyway...
Anyway...
And yes, I did figure out a way to get Active-Active replication going with some EOL sql agent that runs partially on azure and partially on a Windows EC2 that essentially just uses triggers and tracking tables and pick a winner conflict resolution but it's hard as hell
December 29, 2025 at 3:54 PM
And yes, I did figure out a way to get Active-Active replication going with some EOL sql agent that runs partially on azure and partially on a Windows EC2 that essentially just uses triggers and tracking tables and pick a winner conflict resolution but it's hard as hell
Copilot/ChatGTP 4o/Claude 3.5/Claude 4 all choked on this distinction across both the AWS DMS and Microsoft documentation and insisted the impossible was possible
(Azure SQL does support MS.CDC or MS.REPLICATION but only within Azure SQL)
(Azure SQL does support MS.CDC or MS.REPLICATION but only within Azure SQL)
December 29, 2025 at 3:50 PM
Copilot/ChatGTP 4o/Claude 3.5/Claude 4 all choked on this distinction across both the AWS DMS and Microsoft documentation and insisted the impossible was possible
(Azure SQL does support MS.CDC or MS.REPLICATION but only within Azure SQL)
(Azure SQL does support MS.CDC or MS.REPLICATION but only within Azure SQL)
Specifically a lack of transactional logging support for replication/CDC for an active-active database backup out of Azure SQL (of course not, otherwise you wouldn't be locked in)
December 29, 2025 at 3:48 PM
Specifically a lack of transactional logging support for replication/CDC for an active-active database backup out of Azure SQL (of course not, otherwise you wouldn't be locked in)
I'll tell you my absolute 3 months of hell last year was due to being given a task copilot insisted anytime anyone asked it was possible but was actually unsupported due to the differences between SQL Server, Azure SQL Managed SQL Server, and Azure SQL all being 3 related but not identical products
December 29, 2025 at 3:47 PM
I'll tell you my absolute 3 months of hell last year was due to being given a task copilot insisted anytime anyone asked it was possible but was actually unsupported due to the differences between SQL Server, Azure SQL Managed SQL Server, and Azure SQL all being 3 related but not identical products
"Confidently Incorrect" is a term I've heard and adopted myself
"The output of an LLM is language, not truth" is another litany
There's many pitfalls. It will imagine APIs or that APIs do things they're not supported to
"The output of an LLM is language, not truth" is another litany
There's many pitfalls. It will imagine APIs or that APIs do things they're not supported to
December 29, 2025 at 3:45 PM
"Confidently Incorrect" is a term I've heard and adopted myself
"The output of an LLM is language, not truth" is another litany
There's many pitfalls. It will imagine APIs or that APIs do things they're not supported to
"The output of an LLM is language, not truth" is another litany
There's many pitfalls. It will imagine APIs or that APIs do things they're not supported to
My own opinion, ethics not withstanding, is that AI development tools are useful tool in the right context and a nuclear hand grenade in others
It does work best like a sounding board. The Agent and Planning modes in Github Copilot are pretty good at some things and terrible at others
It does work best like a sounding board. The Agent and Planning modes in Github Copilot are pretty good at some things and terrible at others
December 29, 2025 at 3:43 PM
My own opinion, ethics not withstanding, is that AI development tools are useful tool in the right context and a nuclear hand grenade in others
It does work best like a sounding board. The Agent and Planning modes in Github Copilot are pretty good at some things and terrible at others
It does work best like a sounding board. The Agent and Planning modes in Github Copilot are pretty good at some things and terrible at others
Not the kind of mask I hope to see on public transportation especially in cold/flu/covid season
(Also Pertussis/Whooping Cough is on the rise again)
Stay safe, everyone, and @ me if you want a recommendation on masks
(Also Pertussis/Whooping Cough is on the rise again)
Stay safe, everyone, and @ me if you want a recommendation on masks
December 29, 2025 at 3:32 PM
Not the kind of mask I hope to see on public transportation especially in cold/flu/covid season
(Also Pertussis/Whooping Cough is on the rise again)
Stay safe, everyone, and @ me if you want a recommendation on masks
(Also Pertussis/Whooping Cough is on the rise again)
Stay safe, everyone, and @ me if you want a recommendation on masks
Fun Fact
The One Armed Man in The Fugitive was played by the late great Andreas Katsulas, whom I best remember as G'Kar from Babylon 5
The One Armed Man in The Fugitive was played by the late great Andreas Katsulas, whom I best remember as G'Kar from Babylon 5
a man with a leopard print on his face is holding a cup of coffee .
ALT: a man with a leopard print on his face is holding a cup of coffee .
media.tenor.com
December 29, 2025 at 12:05 AM
Fun Fact
The One Armed Man in The Fugitive was played by the late great Andreas Katsulas, whom I best remember as G'Kar from Babylon 5
The One Armed Man in The Fugitive was played by the late great Andreas Katsulas, whom I best remember as G'Kar from Babylon 5