Michael Byrne

I build apps, tools, and software with AI. I also do independent ML/AI research projects and work in harness engineering, evaluation, and training/fine tuning

[Michael Byrne @Mike_Byrne_AI]

Espalier Harness in action!

This is an example of the Write Guard, which blocks writes to protected locations.

Because the hook runs before a tool executes, it puts a major barrier between the agent and your settings, hook scripts, and any rules that an agent could simply move or modify to give itself expanded permissions.

It isn't bulletproof, but most of the mistakes that happen in AI coding aren't:

***Malicious agent goes rogue and deletes everything!

They're much more along the lines of:

*Agent was given approval for a task.
*Agent used that approval to expand its permissions.

The expansion of permissions was never reverted, or worse, was reverted incorrectly.

That allows permissions that were never intended to persist.

The Write Guard is one of the methods
Espalier-Harness uses to make that class of mistake much harder.

[Michael Byrne @Mike_Byrne_AI]

@tec_marco10 Yo! I have that prompt too it slaps!

Ill save everyone some time and just poste it here:

"Gemini, make a video of a futuristic AI themed stock trading dashboard that looks complex with lots of animated graphs and readouts."

[Michael Byrne @Mike_Byrne_AI]

@LeoBuilds_ Ultracode allows the main window to spawn "workflows" which are huge scripted multi agent fanouts.

If you have this on and ask a complex question, it will spawn like 10-20 or even more subagents to swarm a task. This can get very expensive, especially if the subagents are Fable

[Michael Byrne @Mike_Byrne_AI]

@eldridqe I have always wondered how "small" Haiku really is.

Older Haiku models were beating older Llama model in the 405B parameter category in benchmarks.

Haiku may only be "small" compared to like high end frontiers and could be hundreds of billions of parameters.

"Small" is 8B

[Michael Byrne @Mike_Byrne_AI]

Testing edge cases for websites is wild.

*Text turns green if you scroll over it after highlighting another text field in a separate location*

To anyone who has debugged something like that, such a strange conditional issue, WITHOUT AI:

I salute you 🫡🤓

[Michael Byrne @Mike_Byrne_AI]

Don't worry vibe coders, of it can happen to Meta, it can happen to anyone.

Sending a dead link to a frustrated customer in a support chat is about as bad as it gets, but when you have Mag 7 money and economic lock in, who cares right?

[Michael Byrne @Mike_Byrne_AI]

@jahirsheikh8 Day 6: Ask Claude to stop calling you Jason all the time
Day 7: 1500 users

[Michael Byrne @Mike_Byrne_AI]

@pcshipp 3x Max x20 Claude Code
1x Max x200 Codex
1x Google business account
(includes Gemini subscription
1x Supergrok subscription
1x Github copilot base subscription

~900$ a month, before API 🤓

[Michael Byrne @Mike_Byrne_AI]

@AmParth @naval No AGI claims here, just pointing out what distillation is.

[Michael Byrne @Mike_Byrne_AI]

@dark_coderz *Guy with truck to clear out office furniture.

[Michael Byrne @Mike_Byrne_AI]

@itzSaasified An explosion Entrepreneurial spirit, and people profiting from their own self curated value instead of selling it for pennies on the dollar to be a cog in someone else's machine.

[Michael Byrne @Mike_Byrne_AI]

@Saas_addy For sure, there is no method to truly control agents from the same environment they have write permission, "The two step write"
is the issue.

The vast majority of AI errors come from speed and "just for now" edits. This puts a wall between the agent and scope expanding actions.

[Michael Byrne @Mike_Byrne_AI]

Espalier Harness in action!

This is an example of the Write Guard, which blocks writes to protected locations.

Because the hook runs before a tool executes, it puts a major barrier between the agent and your settings, hook scripts, and any rules that an agent could simply move or modify to give itself expanded permissions.

It isn't bulletproof, but most of the mistakes that happen in AI coding aren't:

***Malicious agent goes rogue and deletes everything!

They're much more along the lines of:

*Agent was given approval for a task.
*Agent used that approval to expand its permissions.

The expansion of permissions was never reverted, or worse, was reverted incorrectly.

That allows permissions that were never intended to persist.

The Write Guard is one of the methods
Espalier-Harness uses to make that class of mistake much harder.

[Michael Byrne @Mike_Byrne_AI]

@komal_uk01 I learned the hard way, you don't get to choose.

If you play favorites, your market and opportunity shrinks significantly. You never know when a great opportunity comes by but
"We only use Linux"
kills it.

All your programs should run on all major operating systems.

[Michael Byrne @Mike_Byrne_AI]

@Nishant_N_25 Yes, but because their lack of entry to the "Train a huge proprietary model" game is actually the strategy.

They are purposely absent, and when the time is right, and the kinks are worked out, they can start from a place of knowledge with billions of dollars.

[Michael Byrne @Mike_Byrne_AI]

@samueljmcd Ok, what about this Mr smart man:

If you know so much, how come when I CLEARLY write:

"Make no mistakes"

There are still mistakes?!?

It can't even remember my name and always calls me Jason. AI is dumb.

¿?‽⸮🤓

[Michael Byrne @Mike_Byrne_AI]

@DanielSmidstrup Swich subscriptions :)

[Michael Byrne @Mike_Byrne_AI]

@aslambeg84 Lead AI *insert word*

AI *insert words*

*insert word* AI *insert word*

A *insert word* I

*insert [AI] word*

*AI

Repeat with key terms,
"Vibe, agent, automation"

[Michael Byrne @Mike_Byrne_AI]

I would say Claude changes my mind about a third of the time on issues that I strongly take a position on. Its usually when it points out variables I was unaware of.

I change its mind about a quarter of the time with a viewpoint it cant see or an observation.

Having new information brought into the conversation should change minds.

[Michael Byrne @Mike_Byrne_AI]

@Saas_addy Assuming all 4 are at an equal state, (otherwise the one in the worst shape)

Whatever the user "experiences" the most.

People bounce on a product because they have to put their email in. User experience is everything.

"Make it work like an appliance"
Some visionary somewhere