merklejerk
@merklejerk.com
Burned out blockchain (EVM) engineer, security + rust + AI/ML dilettante, tabletop gamer, compassionate futurist, seasonal misandrist, wasian 🇵🇭 🏳️🌈.
ᛝ⣦⣄▥┱⠇⢤⠜┲╡⣈⢁▣⢆┮⣈⠺╲⡀▝┋▐
ᛝ⣦⣄▥┱⠇⢤⠜┲╡⣈⢁▣⢆┮⣈⠺╲⡀▝┋▐
Where Gemini 3 Pro was very methodical & evidence based, Gemini 3 Flash is intuiting much more of the code "world," but it has VERY good intuition. This is top tier for actual coding workflows, but bad for BS Youtube vibe coding evals, which are very long horizon tasks, so expect them to neg on it 🤡
December 20, 2025 at 11:56 AM
Where Gemini 3 Pro was very methodical & evidence based, Gemini 3 Flash is intuiting much more of the code "world," but it has VERY good intuition. This is top tier for actual coding workflows, but bad for BS Youtube vibe coding evals, which are very long horizon tasks, so expect them to neg on it 🤡
If this holds true then it's possible Google or Meta will dethrone OpenAI. I haven't seen much world model research coming out of OpenAI.
My 2026 prediction is that it will be the year for world models. Not just the spatial world but also digital worlds, code execution worlds, social worlds, etc. This will create a monumental improvement over current models and may finally start to *resemble* AGI.
December 20, 2025 at 11:33 AM
If this holds true then it's possible Google or Meta will dethrone OpenAI. I haven't seen much world model research coming out of OpenAI.
My 2026 prediction is that it will be the year for world models. Not just the spatial world but also digital worlds, code execution worlds, social worlds, etc. This will create a monumental improvement over current models and may finally start to *resemble* AGI.
December 20, 2025 at 11:27 AM
My 2026 prediction is that it will be the year for world models. Not just the spatial world but also digital worlds, code execution worlds, social worlds, etc. This will create a monumental improvement over current models and may finally start to *resemble* AGI.
A year later I still sometimes lay awake at night thinking of that time cardi b gave a speech at the DNC
December 16, 2025 at 1:06 PM
A year later I still sometimes lay awake at night thinking of that time cardi b gave a speech at the DNC
I'm gonna make an alarm app that makes you compare and choose a new individual healthcare plan
December 5, 2025 at 7:27 PM
I'm gonna make an alarm app that makes you compare and choose a new individual healthcare plan
In case you don't understand how absurdly far ahead in the game nano banana pro is...
My prompt: "an infographic illustrating how to identify each member of BTS from one another"
Z image turbo, chatgpt 5, nano banana pro, in that order:
My prompt: "an infographic illustrating how to identify each member of BTS from one another"
Z image turbo, chatgpt 5, nano banana pro, in that order:
November 27, 2025 at 10:59 AM
In case you don't understand how absurdly far ahead in the game nano banana pro is...
My prompt: "an infographic illustrating how to identify each member of BTS from one another"
Z image turbo, chatgpt 5, nano banana pro, in that order:
My prompt: "an infographic illustrating how to identify each member of BTS from one another"
Z image turbo, chatgpt 5, nano banana pro, in that order:
just learned that google's "imagen" is supposed to be pronounced "imagine" and not "image-gen" and I never appreciated such a goated name
November 21, 2025 at 7:33 AM
just learned that google's "imagen" is supposed to be pronounced "imagine" and not "image-gen" and I never appreciated such a goated name
The "AI founder" starter pack pub.towardsai.net/i-reverse-en...
The Real Tech Stack Behind AI Startups: A 200-Company Analysis
Three weeks of network monitoring revealed the truth: 73% of funded AI startups are running $33M valuations on $1,200/month in OpenAI…
pub.towardsai.net
November 8, 2025 at 2:10 PM
The "AI founder" starter pack pub.towardsai.net/i-reverse-en...
"I got a 3D printer for functional prints!"
My functional prints:
My functional prints:
November 2, 2025 at 6:03 PM
"I got a 3D printer for functional prints!"
My functional prints:
My functional prints:
I really find the natural gullibility of humans to either worship or antagonize machinery super fascinating. www.reddit.com/r/Artificial...
From the ArtificialSentience community on Reddit
Explore this post and more from the ArtificialSentience community
www.reddit.com
October 3, 2025 at 11:25 PM
I really find the natural gullibility of humans to either worship or antagonize machinery super fascinating. www.reddit.com/r/Artificial...
Don't be that guy that still believes LLMs alone will achieve true AGI. They only fake causality, spatial awareness, and systemic reasoning. World models might be the key to the first two. HRMs maybe for the last... share.google/Nmbx401g13jO...
Big AI firms pump money into world models as LLM advances slow
World models aim to navigate the physical world by learning from videos, robotic data.
share.google
September 29, 2025 at 4:21 PM
Don't be that guy that still believes LLMs alone will achieve true AGI. They only fake causality, spatial awareness, and systemic reasoning. World models might be the key to the first two. HRMs maybe for the last... share.google/Nmbx401g13jO...
I wonder if ppl complaining about claude code getting nerfed are just for the first time ever entering the last 5-10% of a project, where you have to climb out of the it's-just-a-demo uncanny valley and actually ship to prod. Something all LLMs suck at.
September 11, 2025 at 3:36 AM
I wonder if ppl complaining about claude code getting nerfed are just for the first time ever entering the last 5-10% of a project, where you have to climb out of the it's-just-a-demo uncanny valley and actually ship to prod. Something all LLMs suck at.
hoping the AI bubble bursting saves us from AI the same way the dot com bubble bursting saved us from living in a world where our entire lives were online
August 25, 2025 at 4:24 AM
hoping the AI bubble bursting saves us from AI the same way the dot com bubble bursting saved us from living in a world where our entire lives were online
Next time an LLM wastes 3 hours of your life on a lie, remember that providing convincing hallucinations means the model has successfully learned to generalize a robust solution across the loss landscape instead of simply memorizing training data 🎉. And also you'll never see your $20 again.
August 17, 2025 at 4:34 PM
Next time an LLM wastes 3 hours of your life on a lie, remember that providing convincing hallucinations means the model has successfully learned to generalize a robust solution across the loss landscape instead of simply memorizing training data 🎉. And also you'll never see your $20 again.
Can't wait to see fine-tuned versions of these small models powering 100% in-browser apps share.google/rzrPRjsqsDJ8...
Introducing Gemma 3 270M: The compact model for hyper-efficient AI- Google Developers Blog
Explore Gemma 3 270M, a compact, energy-efficient AI model for task-specific fine-tuning, offering strong instruction-following and production-ready quantization.
share.google
August 17, 2025 at 2:48 PM
Can't wait to see fine-tuned versions of these small models powering 100% in-browser apps share.google/rzrPRjsqsDJ8...
Every leaked AI system prompt has a hint of desperation in it
August 12, 2025 at 2:55 PM
Every leaked AI system prompt has a hint of desperation in it
GPT 5 just called me buddy out of nowhere 🤮. You needy losers ruin everything for the rest of us.
August 10, 2025 at 2:58 AM
GPT 5 just called me buddy out of nowhere 🤮. You needy losers ruin everything for the rest of us.
the only hockey stick graph in AI I'm seeing is the number of vibe coded "professional" websites with exposed personal access tokens
July 31, 2025 at 3:35 AM
the only hockey stick graph in AI I'm seeing is the number of vibe coded "professional" websites with exposed personal access tokens
I am so excited for phishing attacks on dumb agents
July 23, 2025 at 3:48 PM
I am so excited for phishing attacks on dumb agents
Current LLMs are like cooks, but we really need a chef. They can cook really well, but the taste is off. Ensemble scaling techniques are equivalent to just putting more cooks on the same dish, and we know how that idiom ends. We'll know when we have AGI when we see a chef 🧑🍳.
July 22, 2025 at 3:33 AM
Current LLMs are like cooks, but we really need a chef. They can cook really well, but the taste is off. Ensemble scaling techniques are equivalent to just putting more cooks on the same dish, and we know how that idiom ends. We'll know when we have AGI when we see a chef 🧑🍳.
damn you can't get away with nothing these days
July 20, 2025 at 5:14 AM
damn you can't get away with nothing these days
Oh wow you can generate audio overviews right from inside a Gemini canvas 🥰
July 19, 2025 at 6:56 AM
Oh wow you can generate audio overviews right from inside a Gemini canvas 🥰
are FPGAs being used in frontier AI R&D to explore novel primitives? what about APUs? feels like the field may have painted itself into a corner by over-relying on those expensive nvidia matmul processors.
July 16, 2025 at 9:34 PM
are FPGAs being used in frontier AI R&D to explore novel primitives? what about APUs? feels like the field may have painted itself into a corner by over-relying on those expensive nvidia matmul processors.
Before AI slop we had Chuck Tingle
July 16, 2025 at 4:37 PM
Before AI slop we had Chuck Tingle