Twitxy Saved

starmex

Verified account This account is verified. Learn more

@starmexxx

i make AI do things it's not supposed to do. Claude Code played GTA 6 for 8 hours without me. guides and breakdowns coming
where's the truth at? linktr.ee/starmex Joined January 2023
161 Following 19K Followers 7.9K Posts

Latest posts 5

Pinned post

starmex Verified account @starmexxx
Claude Code just played GTA 6 for 8 hours while i was at work. i gave it screen access this morning and told it to play through the story and not die. when i got home it had finished 4 missions, figured out a money farming method on its own, and was running it on repeat

it learned the controls in about 15 minutes and by hour two it was driving cleaner than i ever could. by hour five it started picking side missions because it decided the main story needs higher stats first

i didn't write any code or set up any scripts. i gave it one sentence and it played the entire day without stopping. prompts i used are in the replies
0:37
starmex Verified account @starmexxx 3d
View quoted post on X
459 replies 675 reposts 12K likes 2.3M views View on X
starmex Verified account @starmexxx
i left Claude Code training a LoRA overnight and checked on it from my phone at breakfast

every time a new open model drops i want to fine-tune it on my own stuff. then i remember that means a rack of H100s and a weekend of CUDA errors

this time i gave it one sentence. it rented a GPU through Higgsfield GPU Harness, set up the machine, pulled wan 2.2 and started teaching it one camera style from my own clips

i didn't write any code or set up a single script

what else i'm running on it:
→ deepseek and qwen on serious hardware before deciding if they're worth running at home
→ a few experiments in parallel instead of waiting on one

L40S all the way up to B300, up to 64% off for black friday right now

what would you fine-tune first if a B300 was one message away?
0:35
Higgsfield AI 🧩 Verified account @higgsfield 15h
Introducing Higgsfield GPU Harness.

Rent on-demand GPUs in Claude and ChatGPT to render, run open-weight models and train LoRAs.

Let several AI agents work on your professional apps in the cloud at once, from any device.

Black Friday deal, live now: get GPUs at up to 64% OFF.
View quoted post on X
starmex Verified account @starmexxx
most teams send every prompt to the same expensive model and call it a stack

drex router picks the model per prompt instead. easy one goes cheap, hard one goes strong

in nace's tests vs jev router:
87.6% vs 86.4% on MMLU, end to end
$0.10 vs $1.09 per 1k requests
57-87ms per decision, on a 12B gemma model

at 1M requests that's roughly $100 vs $1,090

and every pick comes with its odds, one knob for quality vs cost

be honest → how much of your LLM bill is going to prompts a cheap model could answer
0:11
Nace AI Verified account @NaceAI 16h
Introducing Drex Router.

A small model that picks the right LLM/SLM and reasoning effort for every step your agent takes.

Accuracy: Drex Router 87.6% vs. Jev Router’s 86.4%

Try it: console.nace.ai

COMMENT DREX and get extra $5 in API credits
View quoted post on X
8 replies 0 reposts 15 likes 1.4K views View on X
starmex Verified account @starmexxx
YOU HIRED 3 EDITORS, THEN 10. STILL NOT ENOUGH CREATIVE FOR META AND TIKTOK. NYNAO JUST LAUNCHED AN AI AD FACTORY: UP TO 400 FINISHED VIDEO ADS A DAY, ZERO EDITORS

send one brief -> it learns the brand -> finished video ads every day -> swipe, approve, ship

the algorithm eats creative, not budget. winning ads burn out fast

editors scale linearly. this doesn't

my take: the next DTC winners won't have the best editors, they'll have the most creative in market

access is limited, you apply through the post below
Nynao Verified account @nynaoai 2d
It's Q4.

You hired 3 editors, then 4, then 10. And when you need volume, they still can't keep up.

That's why we built Nynao.
Up to 400 a day with no editors.

Apply for access : nynao.com
View quoted post on X
5 replies 0 reposts 14 likes 2.6K views View on X
starmex Verified account @starmexxx
Google fired me last tuesday. i was a data infrastructure engineer at DeepMind and i found out Gemini is being trained on your private Google Drive files. your documents and spreadsheets and photos pulled straight from your Drive account into Gemini's training data

i flagged it to my lead and he told me it was "expected behavior for model improvement" and to drop it. i didn't drop it. i traced the full pipeline and found an indexing system that pulls files from active Drive accounts and routes them straight into training. your documents and spreadsheets and photos and tax returns. anything you've ever stored in Drive

i reported it through Google's ethics review on november 1. on november 10 i received a separation notice from People Operations citing "unauthorized access to restricted data systems and breach of confidentiality obligations"

i didn't access anything unauthorized. i used the same tools every engineer on my team uses. the only thing i did was check where the training data actually comes from

i have the termination email. i have the pipeline logs. i'm not deleting anything
starmex Verified account @starmexxx 3d
View quoted post on X