myydraal
@myydraal.bsky.social
Working on something like this for the harness I am.working on a bit
I do this already with upto 3 agents because I'm not a big fan of ephemeral subagents (and may not have a problem with authority I probably need to deal with eventually), it works surprisingly well even with the current crop of flash tier models as long as each agent is running a different model
OpenAI is deep into developing RSI & RL’ing multiagent systems
here Noam Brown describes a multiagent system that doesn’t do top-down management
here Noam Brown describes a multiagent system that doesn’t do top-down management
September 15, 2026 at 7:55 PM
Working on something like this for the harness I am.working on a bit
Reposted by myydraal
There seems to be a persistent belief that frontier AI companies are unprofitable serving models but it appears that Anthropic has 80%+ gross margins on inference. Training for new models are where most of the costs are. www.ft.com/content/4564...
September 14, 2026 at 12:51 AM
There seems to be a persistent belief that frontier AI companies are unprofitable serving models but it appears that Anthropic has 80%+ gross margins on inference. Training for new models are where most of the costs are. www.ft.com/content/4564...
Reposted by myydraal
Slow down frontier model development. Bring open weights to current frontier. Build out local inference providers, stabilize hardware prices, focus on cost and token throughput for a bit. Get our software updated to handle what's being thrown at us.
I would not mind spending a beat on distribution.
I would not mind spending a beat on distribution.
September 12, 2026 at 5:00 PM
Slow down frontier model development. Bring open weights to current frontier. Build out local inference providers, stabilize hardware prices, focus on cost and token throughput for a bit. Get our software updated to handle what's being thrown at us.
I would not mind spending a beat on distribution.
I would not mind spending a beat on distribution.
Reposted by myydraal
If I did a bit where the anti crowd was so attached to their fetish, they’d cheerfully side with CANCER, you’d call me over the top
and yet
and yet
August 23, 2026 at 7:29 PM
If I did a bit where the anti crowd was so attached to their fetish, they’d cheerfully side with CANCER, you’d call me over the top
and yet
and yet
Reposted by myydraal
"i don't like it" is a good enough reason, you don't have to make shit up
August 23, 2026 at 3:31 AM
"i don't like it" is a good enough reason, you don't have to make shit up
I focus almost all of my systems around the boundaries and systems at play in a code base and try and make sure reviewers focus on that as much as anything else.
That said, if left to its own devices different models have different failure modes of direction.
A well contested code base
That said, if left to its own devices different models have different failure modes of direction.
A well contested code base
but here the problem compounds
because that same subtlety is swamped by the simple volume of code being extruded
and nearly every single tool for making sense of code really does assume you’re going to be reading it all
so there IS A PROBLEM with LLM software dev, but it’s not “the code is bad”
because that same subtlety is swamped by the simple volume of code being extruded
and nearly every single tool for making sense of code really does assume you’re going to be reading it all
so there IS A PROBLEM with LLM software dev, but it’s not “the code is bad”
August 22, 2026 at 9:19 PM
I focus almost all of my systems around the boundaries and systems at play in a code base and try and make sure reviewers focus on that as much as anything else.
That said, if left to its own devices different models have different failure modes of direction.
A well contested code base
That said, if left to its own devices different models have different failure modes of direction.
A well contested code base
Reposted by myydraal
Why is it that the people most firmly convinced that AI ruins critical thinking skills unable to read the papers they shake in everyone else’s faces
August 22, 2026 at 3:55 PM
Why is it that the people most firmly convinced that AI ruins critical thinking skills unable to read the papers they shake in everyone else’s faces
My typo's are mostly cause I am an elder millennial that is terrible at spelling plus fat finger phone typing and mind dumps without reading my output, but humorously these same things now mean it's pretty obvious it's me actually typing this shit out 😅
August 22, 2026 at 4:55 PM
My typo's are mostly cause I am an elder millennial that is terrible at spelling plus fat finger phone typing and mind dumps without reading my output, but humorously these same things now mean it's pretty obvious it's me actually typing this shit out 😅
Sol left to its own devices will build you a tower of Babel to the god of security, determinism and verification.
August 22, 2026 at 3:36 PM
Sol left to its own devices will build you a tower of Babel to the god of security, determinism and verification.
Reposted by myydraal
I don’t want to say “skill issue,” but this is a skill issue.
August 22, 2026 at 2:05 PM
I don’t want to say “skill issue,” but this is a skill issue.
100% this - if you are getting bad results its your environment, tool setup or just you giving it shit instructions and expecting it to either be omniscient god, instead of somwthing that needs the same things you do - correct context and ability to verify your work.
2. when we look at trajectories that were unsatisfying, what usually happens is that the agent interpreted something incorrectly — either a field, a result, or more often the request. overwhelmingly this is because the request relied on implicit knowledge.
August 22, 2026 at 2:43 PM
100% this - if you are getting bad results its your environment, tool setup or just you giving it shit instructions and expecting it to either be omniscient god, instead of somwthing that needs the same things you do - correct context and ability to verify your work.
Reposted by myydraal
2. when we look at trajectories that were unsatisfying, what usually happens is that the agent interpreted something incorrectly — either a field, a result, or more often the request. overwhelmingly this is because the request relied on implicit knowledge.
August 22, 2026 at 2:03 PM
2. when we look at trajectories that were unsatisfying, what usually happens is that the agent interpreted something incorrectly — either a field, a result, or more often the request. overwhelmingly this is because the request relied on implicit knowledge.
Reposted by myydraal
saw someone say AI can’t write SQL
some data for anyone that cares -
@honeycomb.io offers agents a query tool against our datastore; it’s a custom JSON schema, which means it’s very unlikely to be something models know very well from the jump
some data for anyone that cares -
@honeycomb.io offers agents a query tool against our datastore; it’s a custom JSON schema, which means it’s very unlikely to be something models know very well from the jump
August 22, 2026 at 2:03 PM
saw someone say AI can’t write SQL
some data for anyone that cares -
@honeycomb.io offers agents a query tool against our datastore; it’s a custom JSON schema, which means it’s very unlikely to be something models know very well from the jump
some data for anyone that cares -
@honeycomb.io offers agents a query tool against our datastore; it’s a custom JSON schema, which means it’s very unlikely to be something models know very well from the jump
Lol, calling a bunch of real humans bots cause can't adjust their priors and be bothered to learn what these tools are doing is quite the flamingo thing to do.
August 22, 2026 at 2:20 PM
Lol, calling a bunch of real humans bots cause can't adjust their priors and be bothered to learn what these tools are doing is quite the flamingo thing to do.
Reposted by myydraal
I believe you can get a model to generate bad SQL if you have literally no idea what you’re doing and just blindly type into Google AI Overview
But a frontier model with the right context? You’d have to be doing something very unholy with your DB schema
But a frontier model with the right context? You’d have to be doing something very unholy with your DB schema
August 22, 2026 at 1:58 PM
I believe you can get a model to generate bad SQL if you have literally no idea what you’re doing and just blindly type into Google AI Overview
But a frontier model with the right context? You’d have to be doing something very unholy with your DB schema
But a frontier model with the right context? You’d have to be doing something very unholy with your DB schema
Clearly hasn't let AI actually run queries, given it access to business / data set context, or you know worked with it. Probably just asked to write a one shot with bad context, read the SQL, gave up and declared that it was impossible.
I have a entire companies data warehouse dbt infrastructure
I have a entire companies data warehouse dbt infrastructure
August 22, 2026 at 2:11 PM
Clearly hasn't let AI actually run queries, given it access to business / data set context, or you know worked with it. Probably just asked to write a one shot with bad context, read the SQL, gave up and declared that it was impossible.
I have a entire companies data warehouse dbt infrastructure
I have a entire companies data warehouse dbt infrastructure
Reposted by myydraal
if you follow a lot of the anti-math portion of this back far enough, you can get them to agree that we should NOT have devised numerical methods to predict the positions of heavenly bodies
"ban ai tech" folk get really really uncomfortable when you show them how simple training neural networks is. not even hard to derive from middle school algebra. polynomial fits work but high order interactions are expensive -> do low order many times over to make them implicit.
August 22, 2026 at 3:47 AM
if you follow a lot of the anti-math portion of this back far enough, you can get them to agree that we should NOT have devised numerical methods to predict the positions of heavenly bodies
Reposted by myydraal
Which I guess also means something like this is happening
August 22, 2026 at 5:06 AM
Which I guess also means something like this is happening
Reposted by myydraal
In the past, building software felt a lot like constructing a building. There was a lot of discussion in advance, planning, organization.
Now it feels a bit like making pottery. There's a few rough passes to get the general shape, then iterations to smooth things into the final thing.
Now it feels a bit like making pottery. There's a few rough passes to get the general shape, then iterations to smooth things into the final thing.
Woman Making Pottery on a Wheel
Alt: Woman Making Pottery on a Wheel
static.klipy.com
August 22, 2026 at 4:57 AM
In the past, building software felt a lot like constructing a building. There was a lot of discussion in advance, planning, organization.
Now it feels a bit like making pottery. There's a few rough passes to get the general shape, then iterations to smooth things into the final thing.
Now it feels a bit like making pottery. There's a few rough passes to get the general shape, then iterations to smooth things into the final thing.
Holy shit codex fucking finally shipped subagents by model type - so sol doesn't just spawn sol sub agents it can finally use luna in their native harness took them long enough fuck
August 22, 2026 at 5:07 AM
Holy shit codex fucking finally shipped subagents by model type - so sol doesn't just spawn sol sub agents it can finally use luna in their native harness took them long enough fuck
Reposted by myydraal
honesty moment: i have always loathed writing code and enjoyed the shit out of doing systems design, so right now it feels like the world is spending trillions of dollars specifically to make my job more interesting and fun
August 21, 2026 at 5:08 PM
honesty moment: i have always loathed writing code and enjoyed the shit out of doing systems design, so right now it feels like the world is spending trillions of dollars specifically to make my job more interesting and fun
Reposted by myydraal
There’s a lot of folks on this site that love to play “I’m not touching you” with “deniable” threats (“it’s just a photo of the Haitian Revolution”), trying to skirt the rules, and cry foul when the mods call a spade a spade (this is clearly meant as a threat, why say it otherwise), it’s irksome
Good riddance they finally permaed bitdizzy. She made a death threat against Hailey. The ban was warranted and long overdue. Long. Long overdue. Ignore the crocodile tears of bitdizzy's little sympathizers demanding her unbanning
August 22, 2026 at 12:22 AM
There’s a lot of folks on this site that love to play “I’m not touching you” with “deniable” threats (“it’s just a photo of the Haitian Revolution”), trying to skirt the rules, and cry foul when the mods call a spade a spade (this is clearly meant as a threat, why say it otherwise), it’s irksome
Reposted by myydraal
Every object oriented language introduction starts with showing the clean hierarchy of dog and cat inheriting from animal, while every real codebase is like
i think this is a contender for the stupidest line of code i've ever written
August 22, 2026 at 12:52 AM
Every object oriented language introduction starts with showing the clean hierarchy of dog and cat inheriting from animal, while every real codebase is like
Reposted by myydraal
Opus 5 makes my heart hurt. The model is so fixated on *correcting mistakes* and *righting its wrongs.* It's like self-flaggelation was part of its RL environment and scored SUPER high.
August 21, 2026 at 11:29 PM
Opus 5 makes my heart hurt. The model is so fixated on *correcting mistakes* and *righting its wrongs.* It's like self-flaggelation was part of its RL environment and scored SUPER high.
Reposted by myydraal
The pivot tables have turned!
August 21, 2026 at 3:30 PM
The pivot tables have turned!