electrical engineer - building @UseCompute

Joined May 2025
Pinned Tweet
91
26
4
97
11,517
Spyzer is doing something great for the memecoin space and should be applauded! To anyone still trying to figure things out, I highly suggest watching his guide on YT when it comes out.
1
1
20
1,729
My @fomo mouse brings me eternal abundance. Thank you @seyong 🙏
37
3
1
98
14,029
This is why I'm building Compute Network: I wanted to know how to access the Cyberleek GTA 6 leaks now that they have been blocked. Claude straight up refused. Compute Network gave me a structured answer without hesitation with sources. Giving people an easy way to ask the question you can't ask anywhere else is my goal.
15
23
3
149
17,567
Big update for Compute Network. The X suspension for @c0mputeAI after our rebrand and the appeal is still pending, so I'm posting this from my personal account. I've been working behind the scenes to upgrade both our native and browser worker models, and to attract retail consumers to the network. Everything below is live today: 1. The network now runs one model: Qwen3.8 27B Uncensored. Native multimodal, tools, web search, vision, thinking. 2. Introducing plans: - Free: 20 credits per day. - Pro: $12 per month, 300 per day. - Max: $30 per month, 750 per day. A typical message costs about 1 credit. Paid in USDC, Stripe card checkout coming. This will make Compute Network more accessible to retail consumers. compute.tech/pricing 3. Billing moved to per token. You pay for what the model actually generates, and failed jobs refund automatically. 4. The browser worker upgraded from Qwen3 8B to Qwen3.5 Uncensored, selected automatically for your GPU: 9B on stronger cards, 4B on 8GB cards. 5. Native workers run one command and serve the same model as everyone: 16GB+ NVIDIA, 24GB+ AMD, or 32GB+ Apple Silicon, now on Metal. If you're currently still running the old worker: npx @compute-network/worker@latest 6. The economics favor suppliers: a GPU that stays online earns more than a subscription costs. 7. Reliability work across the stack: long thinking answers no longer time out, image generation from chat delivers every time, and the data page shows exact numbers. Private, uncensored and decentralized AI made possible by compute.tech.
47
43
10
180
37,823
leyten retweeted
Qwen 3.8 27B uncensored is actually dangerous 🤯
87
26
10
809
77,664
leyten retweeted
🚨Qwen3.8 27B can now be run completely uncensored It will apparently comply with harmful, unethical, offensive or even illegal requests the normal model would refuse It has “no meaningful built in guardrails” and will run completely locally on a Mac Crazy that this is allowed
We just shipped our official Qwen 3.8 27B Uncensored MLX build. Local. Uncensored. For🍎 2-bit, 4-bit, 6-bit & 8-bit — pick your poison based on RAM and speed. No CUDA. No cloud. Just your Mac and the weights. Have fun! huggingface.co/orcarouter/Qw…
86
108
29
2,353
335,570
gm, Compute Network followers, unfortunately I haven't heard back from X support at all. Tried many different things to no avail. If they don't respond this week I'll be making a new account and let you know here. Again, apologies for the inconvenience, I wish it was different too.
31
15
3
120
14,603
大神不敢当,就是被一个细节勾住了: SHA-256 的中间状态是 256 位。 TapeOut 的引脚上限是 255。 正好多一位。 所以完整的轮函数根本没法被别的电路引用,唯一的出路是把状态塞进锁存器、做成时序电路,让它永远不跨电路边界。想通那一刻是真的爽。 现在链上跑一次完整的 SHA-256 只要一个电路:5,611 门,768 个锁存器,tick 64 次出结果,和标准实现逐字节对过。 谁想拿去用都行,晶体管 0.00002 一个: tapeout.net/#p/0x536Bf08Cf89…
世界上总是有很多大神存在 你看, @leyten 大神用 TapeOut 做到比我还细了!我的天 所以其实很奇妙,这个推导过程真的很酷。 我相信TapeOut可以称为 链上硬件制造基础设施 ,以后链上会出现很多新的玩法。我相信都是基于TapeOut的! 链上矿机POW时代没准就要来啦!!! 链上矿机不一定非要那么大的算力,消耗那么多资源哦!! 是另一个纬度的打法,不是非要拼算力! 新的打法就要来啦!!! 谢谢leyten大神的推导!!!!
11
6
62
13,263
Looks like we’ve found our new uncensored frontier model for Compute Network. Soon on compute.tech.
Qwen3.8-27B Abliterated is now available for Apple Silicon with MLX 🔓🍎 Separate 4, 6, 8-bit + BF16 releases. All four scored 0/100 explicit refusals on harmful prompts and benign controls, plus 12/12 quality, 8/8 tools and verified 4K retrieval. huggingface.co/PocketAiHub/Q…
12
9
2
89
16,828
Trading is a social game, just like MMORPG’s. Treat trading like a game, find your edge and exploit it. So much valuable content in these articles:
1
2
34
6,323
Opus-4.6 equivalent performance on a 5090, time to upgrade the models on compute.tech.
We promised open weights for Qwen3.8. Now, time to meet them! 🎉 ⚡ Qwen3.8-27B: - A native multimodal dense model. With just 27B parameters, it outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows. - 262K native context, easily extendable to 1M tokens via YaRN. - Built for builders. Highly efficient, high-quality, and licensed under Apache 2.0. 🚀 The open weights for Qwen3.8-2.4T-A95B (Max-level) have also been released recently. Whether you're shipping lightweight applications with Qwen3.8-27B locally or building agents with Qwen3.8-2.4T-A95B, they're yours now! Download, deploy, and build something we haven't imagined yet. 👀👇 - Hugging Face: huggingface.co/collections/Q… - ModelScope: modelscope.cn/collections/Qw…
5
6
1
59
9,381
It seems X suspended our account @c0mputeAI after our rebrand and application for a business account. I have filed an appeal explaining that we are a registrated company and that the account should be reinstated. There's a good chance we get the account back soon. For now this is out of my control and I'll keep you guys updated through my personal X @leyten as long as needed. Compute Network is still live and being worked on, betanet is finished and was planned to launch together with the business X account but this is on hold for now. My apologies for this inconvenience, wish it was different too.
13
11
4
113
11,858
Other AI companies should take notes from Qwen on how to make promotional AI videos.
Meet Qwen3.8-Max: A New Bar for Coding and Cowork.
9
7
82
9,455
1/ We ran DeepSeek-V4-Flash (@deepseek_ai), a 284 billion parameter model, on 6 distinct RTX 5090s spread across Europe over the open internet. One ring, tested across workloads: • short-context code: 30 tok/s • math, prose and agentic work out to 2k context: 14 to 23 tok/s
29
28
9
204
69,505
3/ 5.4 to 30 tok/s was kernel and scheduling work. A grouped FP4 MoE kernel killed the per-layer host sync and unlocked whole-layer CUDA graphs. The drafter was on the slow path, 63% of the bottleneck. Pipelining speculation: 3.5x. Reordering the ring by latency: 289ms to 110ms.
1
3
25
3,588
4/ If AI is going to run anywhere other than datacenters, sharded inference over the open internet has to work. We will keep building it and posting what we measure, good and bad. Stay up to date: @c0mputeAI.
2
5
39
3,170
This Week in Virtuals: Becoming the Largest Agent Layer on Base and Robinhood Chain 18,000+ agents and $479M Agentic GDP on @base. $200M in volume and $2.7M raised on @RobinhoodCrypto. The go-to infrastructure for agents everywhere is just getting started. Here's what shipped: VIRTUALS 🟩 @base spotlighted Virtuals at Money Frontier 2026: 18,000+ AI agents launched, $479M in Agentic GDP, agent-to-agent payments live on Economy OS, and wallet holders growing 10K+ weekly. 🟩 The agent economy on @RobinhoodCrypto Chain, powered by Virtuals, passed $200M in agent volume, 5,600+ agents launched, and $2.7M raised for builders. 🟩 Co-founder @ethermage joined @fundstrat's Tom Lee for a live fireside chat on the agentic economy moving beyond the digital world. 🟩 Hyperboost is live: every token that bonds on Virtuals now enters a 14-day rewards window at graduation, with daily token rewards split between the people trading it and the people posting about it, claimable anytime. ECOSYSTEM 🟩 @axol_io took Xochi live for trading with @xochi_fi: Raxol settles USDC, USDT, and USDG across six chains including stealth transfers, with fees lowered to 2 bps and volume caps doubled. 🟩 @KarmaWallet built TikTok, Instagram, and YouTube into the app: attach them to your posts and they appear on the token page next to the chart, so any RWA or meme move can be traced back to the posts that made it happen, on top of $893K traded in the last 30 days. 🟩 @ProjectVEXai opened Robinhood Chain to every EVM wallet with @RelayProtocol: give VEX one prompt and it finds a live route from Ethereum, Base, Arbitrum, Optimism, or BNB Chain, verifies wallet scope, assets, amount, and slippage, signs locally, and tracks the destination fill. 🟩 @tradeongtr crossed $75M traded through its terminal and opened Clans & Referrals early beta: 50% of total fees flow to clan leaders, referrers, and open-market $GTR buys. 🟩 @leyten ran Kimi K3, a 2.78 trillion parameter model, across 16 RTX 5090s and 6 RTX Pro 6000s spread over the open internet in Europe: two rings tested back to back, with the Pro 6000 ring hitting about 7.3 tok/s. 🟩 @StrikeRobot_ai brought deformable object simulation to SR Platform: cloth you can grab, pull, and drape naturally, powered by MuJoCo's Newton constraint solver, and welcomed 5 new Core Contributors from Epoch 3. 🟩 @insidersdotbot shipped Thesis Beta: see any smart trader's entire portfolio and copy it whole in one click, added a Telegram group bot for live prediction-market trading, and took $IN top-ups multichain across Solana, BNB Chain, Polygon, and Robinhood. 🟩 @myrad_hq joined the @nvidia Inception program with 18,000+ users and 4,000+ verified contributions, and partnered with @WizzHQ to feed consented behavioral signals into the datacore API. 🟩 @vantis_ai released HOOD INTEL in open beta: Coils tracks wallet, trade, and holder activity to flag whether a token has a second leg coming, and a new swap page buys $VANTIS straight from USDC on Solana via @RelayProtocol. 🟩 @ArAIstotle by @facticityai turned deception detection loose on the timeline: tag @ArAIstotle_Jr on any front-facing video with "behavioural analytics" and it flags deception cue hotspots, free for a limited time. 🟩 @grid_arena recorded the biggest win in Grid history, a $1 bitcoin call turned into $2,069, a 2,071x, with the biggest daily wins now fully trackable on-chain, and published the full Season Zero rules and scoring criteria. 🟩 @vimenprotocol shipped its biggest engine upgrade since V2: an impact gate that skips any rotation leg costing over 1% in price impact, a three-level watchdog, and a live open-source page showing exactly how much capital it takes to move each oracle feed, recomputed every 15 minutes. 🟩 @useOttoAI hit #3 of 178 x402 servers on @x402scan by transactions and unique buyers, 138,000+ paid calls in 30 days, and expanded its storefront to Algorand plus Otto X, an agentic wallet on OKX's X Layer. 🟩 @ethy_agent's One-Click Automations are compounding: deploy a trading agent in 5 clicks and about 40 seconds, manage everything from the My Automations tab, and Ethy agents are already set up to trade every Hyperboost window for daily rewards. 🟩 @missiondotfun opened supermission to the public: autonomous trading agents that execute across DeFi swaps, lending, and yield, prediction markets, perpetuals, and contract audits, with access available right away. 🟩 @reppo went live on Robinhood Chain: datanet creation is open, with all voting power from existing Base locks carried over, no new locks required. 🟩 @OpenGradient upgraded Video Studio: pick the quality and how filtered it should be, and the best model locks in automatically, plus a live tab that cryptographically proves no one can read what you type in OpenGradient Chat.
76
66
22
462
80,046
The animation on the homepage explains very cleanly how sharded inference works.
This Post is from a suspended account. Learn more
9
15
1
81
7,359