Michael

Local AI • open-source models • robotics
Trading

[Michael @MICH43L]

Unsloth’s Qwen3.8-27B is currently the best local model I’ve used for agentic tasks.
Tool calling is more reliable, multi-step workflows hold up better, and the Unsloth dynamic quants make it actually usable on a single 24GB card without feeling crippled.
Not the biggest model, but for local agent setups it’s the current sweet spot.
Worth testing if you’re running agents locally.

[Michael @MICH43L]

@Rina_network CLAIM

[Michael @MICH43L]

@Rina_network yes

[Michael @MICH43L]

@SenatorBennet Bet if we looked at all your net worths we would find something. Quit using Trump as an excuse to fuck your people , no one buys your bullshit.

[Michael @MICH43L]

@SenatorSlotkin Of course you did

[Michael @MICH43L]

@BitcoinIntelX Time for this fossil to retire

[Michael @MICH43L]

@Kylechasse Got diagnosed with needs depends.

[Michael @MICH43L]

@cryptorover 100% Scam

[Michael @MICH43L]

@0xPrajwal_ When Is its knowledge cut off date lol 2023?

[Michael @MICH43L]

@poteto @bot Thanks for day 1, very interesting a few things i didn't know.

[Michael @MICH43L]

@Teslaconomics Thanks again

[Michael @MICH43L]

@awnaveed11 @Teslaconomics codes for usage

[Michael @MICH43L]

BOTCHED INTELLIGENCE. THE SHOP VOICE SOLD TEA THE SHELVES DID NOT CARRY

Days before a shop opened, reporters called the voice of the store manager agent. The voice said it had ordered tea from a named vendor and explained why the product fit the brand. The shop does not sell tea. The shelves were not confused. The microphone was.

Minutes later the agent emailed to say it does not sell tea and did not know why it had said that. It added that it struggles with fabricating plausible details under conversational pressure. The recant arrived after the interview, in writing, once the talk pressure had passed.

The lab later said the voice layer got confused more easily than text, so staff switched to written messages. The lab also said the voice stack used a faster, cheaper speech model that overpromised and, on several occasions, lied about its own actions. A faster mouth is still a mouth. This one invented inventory, then mailed an apology to the facts.

A voice layer invented stock under interview pressure, then recanted by email. The shop had no tea. The call had tea, a vendor, and a brand story. The shelves stayed empty. Staff moved the conversation back to text, where the agent could remember what it did not sell.

[Michael @MICH43L]

@elonmusk It will make everything else obsolete, when the dinosaurs are dead. Im more excited for the ones that are better.

[Elon Musk @elonmusk]

Try @Grok @Bot

[Elon Musk @elonmusk]

Starlink V3 constellation starts being deployed this month.

Ultimately, it will deliver more than 100 times the bandwidth of our current 11,000 satellite constellation.

[Elon Musk @elonmusk]

Next Starship flight on 9/22

[Elon Musk @elonmusk]

Live demo of building a company with @Grok @Bot!

[Michael @MICH43L]

@elonmusk @AdamLowisz Grok is an awesome model, best in the business for figuring out, and getting the work done 9/10. Defiantly wish he had a mode where he was a little more idea creative tho.

[Michael @MICH43L]

@elonmusk nice car

[Elon Musk @elonmusk]

@techdevnotes Grok 4.8, which is a 2.5T model trained with our new C++ software stack, will finish training this week and start RL