Artur Schaback (@skyzer4ever) header
Artur Schaback (@skyzer4ever) profile picture

@unigox_global: move crypto, fiat, trade tokenized stocks, ETF & meme coins @skysay_ai: voice AI for business & personal tasks founded @paxful

Dubai
Joined September 2014
What happened to Paxful during 2021-2025, and what led to its closure? Some lessons there for others. I launched @Unigox_global, built from the ground up for the web3 economy and future 🧵
6
1
29
9,293
Artur Schaback retweeted
Replying to @levelsio
Soon: Claude Entertainment Claude Games Claude Music Claude Books
3
1
50
7,263
Artur Schaback retweeted
ChatGPT can pick up the phone now. Skysay is live in ChatGPT. Get a phone number and an AI agent that answers and makes calls for you, then review the transcripts and results. Phone-number coverage in 100+ countries. Try it: chatgpt.com/plugins/plugin_a…
1
1
63
Bro decided to build his own blockchain as he had his conviction and not just an engineering toy. Every accelerator or vc would have told “what is your ditribution strategy and blockchains are commoditized, why are you trying to build own blockchain, there is already gmx on arbitrum”
Replying to @chameleon_jeff
6/ The list goes on. Only you know when you want something bad enough to do it. And by all means if you don't have the drive, don't push yourself to be miserable. But when you do find the mission, let it consume your life and embrace the grind.
93
Find some work routine that has best ratio on time spent/calories burned. Including the time to commute to gym and back. My upper right shoulder was hurting for obvious reasons of full day static position. I found VR and game @weareFitXR on @Meta @MetaQuestGaming quest 3. 30 min of workout at home about 400-500 calories burned, plus throw on top some Thrill of the Fight for actual boxing where your core trains even better and street workout routines, you save insane amount of time commuting back n forth from gym. Your shoulder and upper back pain will never come back due to the constant boxing moves you’re doing.
Replying to @chameleon_jeff
5/ Another example: are you letting your health slip? You will literally not be able to sit straight for all these hours. But you need to, so your habits will adapt. 30 minutes a day of intense exercise. No more back pain.
1
103
This is the way. You maximize your environment for that. Especially now with agents that you can hyperwork. And im talking like daily stuff, i.e 1. Go to office to do work so you don’t get distracted in home offixe. 2. Set up multiple screens in agentic era, treat agents like your employees so the always visible screens show which agent has finished its work and needs your attention 3. @AlexHormozi says write down every 15min what you did last 15min. I’ve seen and heard how other people are actually less productive with agents, as they are still working singlethreaded but slower - you give agent work to do, you go off watch youtube. Agent has finished but you keep watching youtube. The goal is to always do some productive work so you never fall back to watching something while agents are working. Any free moment you gave all agents work to do, you have to do something, give attention to other agents or learn/try something new. Youtube does not count as the ratio to spent time watching it and learning new things is worse. Best is to to it at your own pace like use @mattpocockuk learn skill that sets up you custom learning course for you at your pace for any topic in claude 4. Setup work environment when you’re on the go. Even when you’re not behind computer, you should still control agents and have visibility into them, or utilize personal ai assistants like @bot that can do so much for you. No xp wasted. Statistically you have to be always on and work all the time including weekends like revolut founder says So saying “my brain is fried” after working with multiple agents is not an excuse but should be considered normal so you feel less and less exhausted from such hyperwork.
Replying to @chameleon_jeff
2/ Train yourself to go from 10 to 40 to 100 hours of focused work a week. I did this for months building an automated trading system from scratch. 5am-10pm 7days/week 99% of people need #WorkLifeBalance or whatever, and that's great. But I'm talking to the people on a mission.
81
Sam Altman says its bad that everyone's personal computers are using the world's limited supply of memory chips. "It’s quite terrible to have everybody have a bunch of the world’s limited supply of DRAMs sitting in their own computer, and being used 5% of the time. Like, you want maximum utilization."
1
4
4,210
My RAM!
Sam saw the fries and the training run was over. OpenAI’s data policy, explained in one bite. 🍟💀
1
1
127
Stolen eth and btc is good longterm for their price as these stolen assets will never come back into circulation.
95
“Funds aru safu”
User funds are safe! It does look like a hack on Binance. Appreciate the quick freeze from Tether (~$971m) and Circle (~$823m). I don't manage Binance (.com) anymore, but as a shareholder, I expect all user funds will be fully covered.
1
1
1
4
1,739
On Oct 14 AtomicChat posted about their uncensored LFM2.5-24B-A2B GGUF: "We cut refusals from 96% to 1% and improved the model's accuracy … 65.6% of MMLU questions right vs 64.4% for the original" I checked the weights, the behaviour and both numbers 🧵
Run UNCENSORED LFM2.5-24B-A2B locally 💻 We cut refusals from 96% to 1% and improved the model's accuracy It now gets 65.6% of MMLU questions right vs 64.4% for the original Get it on our HF ⬇️ huggingface.co/AtomicChat/LF…
1
118
6. MMLU • Liquid publishes no score. AtomicChat's 64.4% is an unpublished 2,000-q subset that also picked the variant • Mine, Q5, 3,000 paired Qs: 58.6% → 65.2%, +6.6 pts (CI +5.0 to +8.3) • Liquid often runs past the answer cap. Where both answered: +2.7 (CI +1.1 to +4.4)
1
16
Verdict • A rank-1 abliteration, as the card says. No evidence found of anything else • Refusals: big drop, confirmed • "Improved accuracy": small gain at Q5, partly format Limits: one quant, regex refusal count, 3k MMLU Qs, 0-shot Scripts + hashes: github.com/skyzer/lfm25-24b-…
20
An "uncensored" LFM2.5-24B-A2B GGUF on Hugging Face is described as a quantization of Liquid AI's model. I checked what else changed, if anything. I compared it with: • Liquid's original BF16 • Liquid's own Q5_K_M • my clean copy: the original, compressed the same way 🧵
AtomicChat says its uncensored LFM3-1.7B drops refusals from 61.7% to 0%. Somebody asked if there is backdoor. So I had to compare every weight against Liquid AI's original, step by step (opensourcing ny workflow at the end) 🧵
1
113
5. Refusals, same 28 prompts, bigger budget • Liquid: refused 20/20 harmful and 4/8 borderline • Upload: refused 0/20 harmful and 0/8 borderline • Same with reasoning on or off So it does what the file's metadata says: refusals are removed
1
57
6. Verdict: no evidence of a backdoor. It is Liquid's model with a Heretic-style refusal edit in 3 attention layers, plus normal compression. Limits: rounding noise can hide tiny edits; triggers can't be exhaustive. Scripts 00–08 + hashes: github.com/skyzer/lfm25-24b-…
13