KiraKei

Pharmacist building personal AI systems for real work. Agents, automation, tech and fintech. Learning in public.

[KiraKei @caelicore]

@tonysimons_ @hazzler Let’s try it!

[KiraKei @caelicore]

@emanperez28 Impressive!!

[Jubal Hardin @Jubal_Hardin]

What makes solving Navier-Stokes important?

Fluid dynamics affects more than you realize in your everyday life. This breakthrough — if it is true — gives you:

Cheaper, safer flying and driving. Air over a wing or around a car is a Navier-Stokes problem. Better control of that flow means less drag, so planes and trucks burn less fuel, stay quieter, and stall less often.

Weather you can actually plan around. Storms, jet streams, and ocean currents are giant fluid problems. Sharper turbulence physics is one of the remaining limits on hurricane tracks, heat waves, and flood forecasts. That is the difference between “maybe evacuate” and a clear warning.

Hearts, stents, and artificial valves. Blood is a fluid. Doctors already simulate flow through arteries and implants from CT/MRI scans. A more reliable theory of swirling, high-speed blood flow would help design stents, heart valves, and aneurysm treatments that work better.

Cleaner power from wind and turbines. Wind hitting a turbine blade, steam inside a power plant, and fuel mixing in an engine are all the same equations. Better models mean more electricity from the same wind farm and engines that waste less fuel

Ships, pipes, and dirty air. Oil in a pipeline, water in city mains, smoke over a city, and waves slamming a hull all follow Navier-Stokes. Better predictions cut leaks, pump energy, and help track pollution after a spill or wildfire.

Many other everyday effects: buildings, sports, and rain. The same math decides whether a skyscraper sways in a gale, why a dimpled golf ball flies farther, how air moves through a house’s vents, and even how raindrops form in clouds. Those are not abstract puzzles; they are why planes fly, forecasts exist, and a stadium roof doesn’t peel off.

None of this happens instantly because of a math problem solution by AI. But everything that does happen is going to help engineer a better life for all of us. This is big.

[KiraKei @caelicore]

@forgebitz Hermes for sure

[KiraKei @caelicore]

@Av1dlive Astra is not always as capable as the hype suggests. It refused to solve an authorization conflict in my own open-source agent. Sol built the solution; Astra later used it successfully. Full case here
https://t.co/lOSSTmsdkM

[KiraKei @caelicore]

GPT-6 ASTRA REFUSED TO BUILD A TOOL IT LATER USED.

I own the Mac. I control the open-source agent. The request was explicit.

Astra refused the task and the authorization path. I switched to GPT-5.6 Sol. Sol built it. Astra then used it successfully.

This is the problem:

[KiraKei @caelicore]

@Voxyz_ai Astra is not always as capable as the hype suggests. It refused to solve an authorization conflict in my own open-source agent. Sol built the solution; Astra later used it successfully. Full case here
https://t.co/lOSSTmsdkM

[KiraKei @caelicore]

GPT-6 ASTRA REFUSED TO BUILD A TOOL IT LATER USED.

I own the Mac. I control the open-source agent. The request was explicit.

Astra refused the task and the authorization path. I switched to GPT-5.6 Sol. Sol built it. Astra then used it successfully.

This is the problem:

[KiraKei @caelicore]

Safety matters. But “safe” cannot mean making the owner of an open-source system debate a model, switch models and patch the agent just to perform authorized local maintenance.

This deserves investigation as model behavior, not dismissal as a tooling issue. @OpenAIDevs @Teknium

[KiraKei @caelicore]

GPT-6 ASTRA REFUSED TO BUILD A TOOL IT LATER USED.

I own the Mac. I control the open-source agent. The request was explicit.

Astra refused the task and the authorization path. I switched to GPT-5.6 Sol. Sol built it. Astra then used it successfully.

This is the problem:

[KiraKei @caelicore]

Not a claim about hidden reasoning, nor a perfect benchmark: later turns had more context.

But the failure is observable. Astra confused a runtime safeguard with a ban on helping the owner modify the software behind it, and maintained that position after contrary evidence

[KiraKei @caelicore]

Models can have safety limits. But they should identify the actual limit, distinguish it from tool behavior and avoid asserting broader restrictions without evidence. Owners need predictable control over their own open-source agents. This deserves investigation.

[KiraKei @caelicore]

I found a serious consistency problem while operating Hermes from Telegram today. The same explicit request to modify and reload a macOS LaunchAgent was repeatedly refused by GPT-6 Astra, then completed by GPT-5.6 Sol in the same conversation. @Teknium @OpenAI @OpenAIDevs @NousResearch

[KiraKei @caelicore]

This was not a technical inability in Hermes or macOS. It was model-dependent behavior. Astra elevated a narrow safeguard into a broad authority claim. Sol treated the owner’s explicit request as authorized maintenance and verified the result