abhinav

I really love AI and Biology

[abhinav @beyondwudan]

@DavidHein238521 @theo No I didn't get why people used t3 because harness is something you choose very selectively for your work. And I didn't really expect Theo to reply...

[abhinav @beyondwudan]

@theo I hear you T

[abhinav @beyondwudan]

@Da7_Tech I stay respectful and so does all of my hermes agents all running on GPT 5.6 Sol.

[abhinav @beyondwudan]

A lot of things in mind but here 2 which really contributes to 2 of my favorite communities,

1 - MorphoStack, an open-source tool that turns 3D microscope images of cells and vesicles into reproducible measurements. It flags uncertainty instead of guessing and links every result to the raw data. I need to ship the validated Windows app, public benchmark, and paper ready release for labs around the world without expensive software or GPUs.

2 - IRLBench. Most AI benchmarks test models on coding puzzles. IRL Bench tests something harder to measure: whether a model behaves sensibly in messy real situations, where it has to decide under uncertainty, notice it was wrong, recover, and know when to hand control back to a person.
You write a set of realistic scenarios, run several AI models through each one, and publish how they did and where they failed. And some of those scenarios come from real genomics work!

[abhinav @beyondwudan]

HTMLs and Prompt - https://t.co/EfqXkZq7Gq

[abhinav @beyondwudan]

Before Gemini 3.8 Flash releases in a few hours, just wanted to take a look at how far @GoogleDeepMind has moved the needle in the flash series...

It's really impressive for a flash model to be this good. Can't wait for 3.8 especially how much rumors are hyping it up!

Prompt below...

[abhinav @beyondwudan]

@Tap4Tchnologies HTML with @threejs

[abhinav @beyondwudan]

Fable 5.1 Max made a cinematic scene of Earthians landing on an alien forest...

It absolutely killed, this model has some good taste sir!

Claude is so back!

[abhinav @beyondwudan]

Another Fable 5.1 Max output for the same scenario.

Amazed by both outputs. This model is gonna take humanity beyond possibilities...🚀

[abhinav @beyondwudan]

Fable 5.1 Max made a cinematic scene of Earthians landing on an alien forest...

It absolutely killed, this model has some good taste sir!

Claude is so back!

[abhinav @beyondwudan]

A new version of Qwen 3.8 Max was just deployed...

Fable 5.1 has competition now.

[abhinav @beyondwudan]

@AbdoKerdawy dont worry G its gonna get fixed theyre getting sued af

[abhinav @beyondwudan]

Something changed with Claude Code limits, back then it was using 12.5% per 5h session...

[abhinav @beyondwudan]

@AbdoKerdawy 😂😂

[abhinav @beyondwudan]

I really want to do Harness battles...!

[Knowix @knowixbuilds]

GPT-5.6 SOL with DeepSeek Harness is insane at creating voxel dioramas.

it one shotted a playable voxel battlefield 👇

That includes:
- trenches, craters, ruins, coastline and railway
- 34 soldiers + 17 animated land, air and naval vehicles
- destructible buildings, artillery and airstrikes
- tactical radar, cinematic camera, night mode and audio

every operation worked. zero runtime exceptions.

[abhinav @beyondwudan]

Imagine OpenAI pulling out on Hermes as a provider now...💀

[abhinav @beyondwudan]

@Da7_Tech Hey Da7em, I asked you in a previous post about your favorite memory plugin in hermes, I'm still curious... I use honcho but I'm thinking of self hosting it because I burned through the credits very quickly

[abhinav @beyondwudan]

https://t.co/fs1X93jXH2

[abhinav @beyondwudan]

YES HY4 PREVIEW IS CRAZYY!

I'll give Hy4 Preview the sunrise.

But Gemini 3.7 flash high gets the lead on orbital salvage.

HTMLs and Prompts below -

[abhinav @beyondwudan]

I am a huge supporter of open models, but bro, MiniMax M3 and Muse Spark 1.2 do not come close to Luna Max here.

Then I added Sol xhigh because why not. It finished the repair, unfolded both arrays, and beat Luna on the full 32-second sequence. I did not expect that.