Shailesh

Building in Stealth | Independent AI Researcher | Victim of AI Psychosis

[Shailesh @0xThoughtVector]

No more gentle singularity, anon.

[Shailesh @0xThoughtVector]

@Safinazelhadry @curatedbykamila @yunbin_bae @luisgnet @farisamagazieva @OpenAIDevs @founders_cafe @quasa0 @luisaalvarez95 bcz the tabs triggered me

[Shailesh @0xThoughtVector]

@Safinazelhadry @curatedbykamila @yunbin_bae @luisgnet @farisamagazieva @OpenAIDevs @founders_cafe @quasa0 @luisaalvarez95 I was really struggling to look at your video ngl

[Shailesh @0xThoughtVector]

@MillionInt In hindi it means Garbage.

[Shailesh @0xThoughtVector]

@deepfates and often the resets come like one day before or after a cycle ends.

[Shailesh @0xThoughtVector]

@Ananthr27104587 $100 Pro sub. I do extra high for co-ordinator thread but it strategizes and gets other threads to do all the work. They're all 5.5 High.
Use 5.4 High and 5.4 mini for some things as well.

[Shailesh @0xThoughtVector]

I have seen humans who are paid to build svg graphics make shittier svg graphics. This is beyond great: it's perfect. Wow!

[Chris @ChrissGPT]

Claude Mythos API checkpoint claude-oceanus-v1-p, had leaked last-night and it looks like they got the cost below 100$ per million token (output)

Svg outputs are looking really nice

Cc: @chetaslua

[Shailesh @0xThoughtVector]

@petergostev what's the update today? I updated but see no apparent difference

[Shailesh @0xThoughtVector]

umm... this is perfect. what the hell

[Chetaslua @chetaslua]

🚨 Here is your First Output of Mythos class model from anthropic

claude-oceanus-v1-p ( it was available for red teamers and available on chinese forum)

Linux(.)do is a Chinese forum

For more information on the context check the profile @scaling01 @synthwavedd & follow them

[Shailesh @0xThoughtVector]

This is the multiple with the current models. 8x productivity gains. And it probably includes a 1.5-3x multiplier for the amount of effort the humans are putting into the work compared to 2021.

I imagine a smaller team with even less bureacracy could go to maybe 10-20x but that's that for now.

No 100x.

[Anthropic @AnthropicAI]

Today, Anthropic engineers on average ship 8x as much code per quarter as they did compared to 2021-2025.

[Shailesh @0xThoughtVector]

A conductor thread created this thread and chose the 'gpt-5.5-codex' model by mistake.

Codex doesn't deny its existance, only that it can't be accessed by a ChatGPT account. Presumably this is the better coding and AI Research model used internally at OpenAI.

[Shailesh @0xThoughtVector]

@petergyang everythign uses Agents.md except Claude. OpenAI guys tried to make Agents.md a standard and everyone complied except Anthropic

[Shailesh @0xThoughtVector]

An Architecture can have the ability to learn enough underlying structure that when given the write inputs and access to feedback, it leads to emergence of consciousness.

Seems like the transformer arch is good enough for that but he jury's still out.

[kasra @kasratweets]

this is an interesting point in the new ted chiang piece – no one really claims that alphafold is conscious, or that sora or midjourney or dall-e are conscious

[Shailesh @0xThoughtVector]

@reach_vb Bro someone changed the highlight colour when text is selected and it is EXACTLY the same as the background colour in the Vercel Dark Mode Theme. Abysmal visibility in other themes as well. Please fix.

[Shailesh @0xThoughtVector]

@ThePrimeagen Didn't you post 'days since AGI' a few months ago? Something, which I based on you having access to big lab insiders and thus updated my priors which led me to start my company.

What the fuck was that?

[Shailesh @0xThoughtVector]

@boazbaraktcs @tszzl A more powerful conscious entity is a different ballgame

[Shailesh @0xThoughtVector]

https://t.co/ZzJy9om3HF

[OpenAI Developers @OpenAIDevs]

Look closely. There’s more in the Showcase.

[Shailesh @0xThoughtVector]

@thsottiaux @reach_vb encountering a bug in codex where text selections don't get highlighted even though 'Add to chat' and 'Ask in side chat' are visible.

Happens when a thread completes a few turns.

[Shailesh @0xThoughtVector]

@teodorio Anthropic is a really, really shit service provider too.

Abysmal limits, high costs of models, lack of communication around limit reductions/fluctuations, model quality unreliability.

[Shailesh @0xThoughtVector]

Holy shit. This is fucking crazy good.

[OpenAI @OpenAI]

It's time to fly.