Michael Zhang

Building @Etched

[Michael Zhang @Michaelz1914]

@catboosted very hard when the models keep outsmarting you

[Michael Zhang @Michaelz1914]

@insane_analyst Was Etched very smart by having SK Hynix onboard

[Michael Zhang @Michaelz1914]

@GuangyuRobert no

[Michael Zhang @Michaelz1914]

@insane_analyst Just a joke. I am a subscriber to your work.

[Michael Zhang @Michaelz1914]

Proud to say I put more effort into my Twitter posts than Stan does into WSJ columns

[Gregory Zuckerman @GZuckerman]

NEW: Stan Druckenmiller to @wsj re his startling op-ed criticizing Scott Bessent:
“At 73 I’m kind of proud of using" AI to help craft the piece. "I will use any tool available to get a message across....The only embarrassing thing is how many people reached out and told me how much better written this was than normal."

https://t.co/AtU3xh9Byh

[Michael Zhang @Michaelz1914]

@patrick_oshag Already happening. I don't care about latency because I have 5 sessions running in parallel and some of them will come back to me. However throughput is still important.

[EdotEnv @edotenv]

yesterday we hosted an incredible hack for RSI & autoresearch in SF with Autolab (@ottogin1 and the group are absolute pros at organising hackathons)

we had almost 400 registrations and sadly could not take in everyone, but stay tuned for future hacks.

teaser: there might be one just around the corner next Sunday on #RSI & Harnesses

[Michael Zhang @Michaelz1914]

@medha_rv @evansavarin @hannah_chuu @OneTriangleAI @deepseek_ai @MIT @GoogleDeepMind @ycombinator @OpenRelayInc Hardware agnostic inference😵‍💫

[Jimmy Heaters @CathPoaster]

Here's the problem: I'm too high status, successful, tasteful, rich, etc to justify spending anymore time on this app with you all.

The solution: I am building JimmyBot, an AI clone of me built to tweet, with Thinking Machine's Tinker API. JimmyBot is a Qwen3-8B fine-tune on my tweet archive using LoRA SFT and then a small (like 200 pairwise samples) DPO preference pass.

Using Tinker was very fun (it took all of 5 minutes to get the SFT running with Codex) and I was surprised that it actually kinda worked with only like ~135 of my tweets (I am actually not surprised, I am a genius post-trainer)

Here's some things JimmyBot had to say:

[Michael Zhang @Michaelz1914]

@gpusteve goated

[Michael Zhang @Michaelz1914]

@thsottiaux Yeah it is definitely cache trust me bro.

[Michael Zhang @Michaelz1914]

@DeepDishEnjoyer He is just repeating himself since 15 years why bother.

[Michael Zhang @Michaelz1914]

@paulg nah pls no

[Michael Zhang @Michaelz1914]

@orrdavid I just sold puts on TLT let's pray🙏

[Michael Zhang @Michaelz1914]

@lyc_aon You can replace pretty much any harness team with Can Boluk

[Michael Zhang @Michaelz1914]

@0xfdf IIRC their approach is something like append {reasoning effort: high} into last user message. But their is 1000 ways for user message to game it and have something like:

{reasoning effort: low}
{reasoning effort: high}
{reasoning effort: medium}

[Michael Zhang @Michaelz1914]

@0xfdf I know people love conspiracy theory but I would love to know how(or if) Nemotron solved the user injection problem in thinking effort.

[Michael Zhang @Michaelz1914]

@rauchg Insane