←
Alexey Fateev7,380 posts
Alexey Fateev avatar✓

Alexey Fateev ✓

@superalesha

how far can 4x RTX 3090s go?

◷ Joined January 2026▧ 697 Photos and videos
Alexey Fateev avatar
Alexey Fateev@superalesha·3h

🎉 Introducing Redcast! Open-source Raycast. Raycast put AI on credits and locked BYOK, your ChatGPT and Claude subs, custom endpoints, and local Ollama behind a paid plan. I got tired of paying to use my own keys, so I replaced Raycast. It's my daily launcher now. It's native SwiftUI and AppKit under 100 MB with no Electron, and I brought over the daily tools I actually use in Raycast. Real Raycast Store extensions run natively with AI.ask wired to your own model, alongside clipboard history with OCR, snippets, quicklinks, 35 window actions, layouts, notes, calendar, calculator, and .rayconfig import. For AI, you bring whatever you want. Connect OpenAI, Anthropic, Gemini, OpenRouter, a local server, or Sign in with ChatGPT to run on your Codex plan without an API key. Agents live right inside the launcher. Run folder-aware agents in the background with @.agent or on schedules with spend caps. You can also run Claude Code, Codex, Gemini CLI, OpenCode, Goose, and Hermes over ACP and resume any past session in 1 key. Double-tap right Command for Screen Awareness to drop the front window straight into AI Chat, or dictate into any field using on-device models like Parakeet and Qwen3-ASR or your own cloud key. Ask and Edit by Voice work from a held modifier key too. There's no account or subscription and zero telemetry, for $0. Forked from Tinycast under AGPL-3.0 and now maintained on its own path for Apple silicon on macOS 26+, signed and notarized by Apple with automatic updates. brew install --cask alesha-pro/redcast/redcast redcast.io github.com/alesha-pro/redcas…

0:47
Alexey Fateev avatar
Alexey Fateev@superalesha·2h

"JUST IN" my ass. What the fuck is breaking about a filing from Nov 10? Anthropic put Nvidia's $10B share purchase in its amended S-1 8 days ago, and Reuters reported the talks back on Sept 11. Nothing new happened today. Sept 11: Reuters reports Nvidia is in talks to put up to $10B into Anthropic's IPO. Nov 10: Anthropic's S-1 Amendment No. 2 lists a $26B Class N placement across 7 chip and cloud companies at the IPO price. Google and Nvidia are $10B each, AMD is $2.5B, Broadcom, Micron, and Microsoft are $1B each, TSMC Partners is $500M. Nov 17: Nvidia's interim CEO repeats that same $10B IPO plan on the Q3 earnings call, 2 days before Anthropic lists on Nasdaq on Nov 19. Why are people acting shocked in the replies when this has been public in Anthropic's own IPO docs for over a week? Slapping "JUST IN" on an 8-day-old SEC filing is dirty clickbait.

Post attachmentPost attachment
Alexey Fateev avatar
Alexey Fateev@superalesha·8h

HY4 FLASH WEIGHTS JUST DROPPED!! Tencent opened Hy4 flash, FP8, base, and day-1 llama.cpp patches. 220B total, 1M context, and 4-bit fits on 4x 3090 with ZERO expert offload: 63.5B is the Engram table in RAM or NVMe mmap, so only 155.5B goes on GPU (96 GB VRAM or a 128 GB Mac). DeepSeek V4.1 Flash 1017 drags ~290 GB of FP4 experts before you even count Engram, so prefill crawls at home. Hy4 flash keeps every expert in VRAM. Hybrid 1:3 sparse/sliding-window attention and LatentMoE shrink KV cache to 1/6 of Hy4 and 1/24 of Hy3, and the built-in DFly2 drafter (0.85B) reuses that sparse KV cache with no extra KV memory. AngelSlim already posted Metal and CUDA patches pinning Engram to CPU mmap. Q4_K_M hits 48.1 tok/s decode (629.8 prefill) at 4K on an M5 Max 128 GB while the 126 to 157 GiB GGUFs finish uploading. Downloading the weights right now, gonna be cooking on my 4x 3090 rig soon.

Post attachment
Alexey Fateev avatar
Alexey Fateev@superalesha·13h

US labs charge $200 a month and still throttle paying users mid-session. DeepSeek runs a $0 app with no ads for ~200M daily users, clears 7-day-old idle KV caches, and people act shocked. Meanwhile their API is untouched, V4.1 Flash still leads OpenRouter at 59.2T tokens a week, and you can run the weights yourself. x.com/aleabitoreddit/status/…

Post attachmentPost attachment
Alexey Fateev avatar
Alexey Fateev@superalesha·16h

Europe burned €7M of public GH200 and H100 compute across 3 supercomputers to ship a 7B model with 1B active parameters and a 32K context window in late 2026. Apache 2.0 weights are great. Calling €7M for 7B-1B plus AI Act compliance paperwork a breakthrough is wild. x.com/olesovhcom/status/2122…

Post attachment
Alexey Fateev avatar
Alexey Fateev@superalesha·20h

Bring Jensen Huang back as NVIDIA CEO. Jensen's statement today took real guts. Most execs hide behind lawyers and blame everyone else when things go south, while he owned his mistakes flat out and offered to step back in if the board asks. Seeing NVIDIA employees and partners already rally behind him shows how much loyalty he earned. You can't replace 33 years of conviction with a committee. Almost every developer training a model or running open weights on their own GPU relies on what Jensen and his team built from scratch. When politicians started pushing to lock AI behind federal licenses and shut down open models, he actually stood up for independent builders instead of playing it safe. Everyone makes mistakes in their personal lives, and owning them in public isn't easy. Pushing out the founder who turned NVIDIA into the engine of modern AI right as this race gets tougher would hurt NVIDIA and everyone depending on it. Keep your head up, Jensen. The people building the future want you back leading NVIDIA.

Alexey Fateev avatar
Alexey Fateev@superalesha·22h

They finally stopped pretending this is about safety and celebrated the monopoly on camera. Cut Nvidia off from open weights and hand Anthropic the market right before its IPO. Funny how every "national security" panic ends in the same place.

2:57
Alexey Fateev avatar
Alexey Fateev@superalesha·Nov 17

Everyone swore memory shortages and infinite AI CapEx would last forever. That bullshit just cracked in 4 places in 48 hours. Spot 512Gb NAND wafer prices dropped 20% since late October to $14.30 (down 12% in a single week and 32% QoQ), per BofA. PC and phone makers finally walked away from inflated prices, older fabs (20 to 30% of supply) are dumping inventory, and BofA expects another 5 to 10% drop in Q1. Morgan Stanley downgraded a batch of Taiwanese memory names (Winbond, Macronix, Phison, PSMC) and cut its Nanya target, warning that cloud CapEx growth is cooling off. SemiAnalysis slashed its 2027 to 2030 global AI accelerator CapEx forecast by $1.6 trillion (2028 cut 28% to $758B, 2030 cut 40% to $923B). Nvidia takes the hardest hit: revenue forecasts down 22 to 66%, 2028 HBM demand cut 30%, and estimated Feynman Ultra prices chopped 44% from $175K to $98K. And banks are unloading AI data center debt. FT reports ING and SocGen put up to $800M of Oracle-linked loans up for sale. Oracle already declared force majeure on Project Jupiter in New Mexico over power permit delays, so those loans are pricing around 85 cents on the dollar. Some big lenders also walked out of Anthropic's $63B debt deal. You can't print grid permits, and PC and phone makers are already cutting memory specs instead of paying inflated prices. When banks start dumping Oracle cluster loans at a 15% discount and even SemiAnalysis cuts next-gen Nvidia GPU prices by 44%, the "shortage forever" story is dead.

Alexey Fateev avatar
Alexey Fateev@superalesha·Nov 17

OpenAI couldn't even get GPT-6 Astra to follow basic PCB factory tolerances for a desktop voice toy. Meanwhile a Chinese startup called CosmosMind already taped out an AI-designed photonics chip in a real Fab, AND shipped working DeepSeek V4.1 kernels for Ascend 910C. I opened their GitHub today. PyPTO-FlashDSA has 0 stars. Nobody in Western AI Twitter is even looking at it. In OpenAI's own November 3 post about their TamaGPT gadget, they admitted GPT-6 Astra failed to consistently respect manufacturing limits and assembly rules on a simple circuit board, and it scored 37% on RWE-Bench. CosmosMind plugged their MetaRSI agents straight into Fab process tolerances and real NPU racks instead. The optical chip came back from the Fab with 31 SEM scans confirming line-width variation under 5 nm (half their 10 nm limit), hitting -0.66 dB coupling loss with 0 backup test arrays on the wafer. Then the agents wired 3 variants of 1.6 Tbps (8x200G) optical links down to 1.46 pJ/bit wall-to-wall. The GPU side is even wilder. DeepSeek's official Ascend C port for V4.1 doesn't even compile on Ascend 910C (it targets the newer 950). CosmosMind's agents wrote PyPTO-FlashDSA from scratch on real 910C cards and beat the official fp32 reference decode kernel (54.45 µs vs 56.49 µs). They also took MiniMax H3 (33B video+audio), which lost 71% of its speed when ported off CUDA, fused ~8,000 kernel ops into 1, and pushed it to 450% of MiniMax's own release. Some of their Qwen3 patches already merged upstream into Huawei's pypto-lib. Domestic chips passed 41% of China's AI accelerator shipments in 2025. While OpenAI struggles to route a simple voice toy without human babysitting, Chinese teams are using agents to fix their own compiler stack and pull working optical silicon out of a Fab.

Alexey Fateev avatar
Alexey Fateev@superalesha·Nov 17

151 tok/s on 2 RTX 3090s, plus a 224B multimodal MoE fitting in 96 GB of VRAM with NO PCIe offload. Qwen4 Flash and HY4 Flash are the ONLY 2 models I'm waiting for right now. Both labs promised open weights, and both dump tens of billions of params onto RAM or NVMe. Alibaba hasn't posted the exact size of Qwen4 Flash yet, only that smaller Qwen4 open weights are next. Still, I already run its qwen4_exp preview every day, Qwen3.8-Flash-Next (125B weights + 51B N-gram table, 6B active). In Strata with Unsloth UD-IQ4_XS, DFlash2 and the 29 GB N-gram shard on NVMe, I get 110 tok/s on code on 1 RTX 3090 and 151 tok/s on 2 cards at 262K context. Qwen4-Max just cut token usage on Terminal-Bench 4.0 by ~49% over 3.8, and llama.cpp, Strata, vLLM and SGLang already have qwen4_exp kernels ready. If Qwen4 Flash stays near that size, 1 or 2 gaming GPUs get a massive upgrade on day 1. HY4 Flash is 224B total, 12B active, 1M context and natively multimodal. 64B of that 224B is Engram lookup memory that stays in RAM or NVMe without hurting decode. That leaves 160B of real weights, around 80 GB at 4-bit, so all of it fits in 96 GB of VRAM on 4x 3090 or unified memory on a 128 GB Mac (~64 GB at 3.2 bpw on 3x 3090). DeepSeek V4.1 Flash 1017 has ~290 GB of FP4 experts before Engram, so at home it streams experts over PCIe and prefill crawls. HY4 Flash needs 0 expert offload, and its KV cache is 1/6 of HY4. Even though its long-horizon coding score is low (19.2 on Terminal-Bench 4.0 vs 44.4 for full HY4), everyday agent tasks match DeepSeek V4.1 Flash at 50% fewer output tokens. I plan to use it as a daily local agent for the tasks I don't want to hand over to cloud models. Qwen4 Flash covers 1 to 2 cards at 100+ tok/s, and HY4 Flash fills 4x 3090 in pure VRAM with 0 offload.

Alexey Fateev avatar
Alexey Fateev@superalesha·Nov 16

This light is driving me crazy. I gave Gemini 4 Carbon a shot through @factory and honestly, it’s a good model. This scene came out of 1 prompt with zero fixes. Every candle is a real light. Drag one and the shadows swing, the wine throws red caustics across the cloth, and the glints on the glass are insane. 1 HTML file, no assets, all of it built in code. YuE2 music on my RTX 3090

1:02
Alexey Fateev avatar
Alexey Fateev@superalesha·Nov 16

Ok, I got way too obsessed with light this time. I gave 4 models the same prompt. A glass vase built from Gaussian splats, light you drag with the mouse, real refraction, caustics and shadows from EVERY object. Kimi K3.1 Pro is my favorite. On light it's on par with GPT 6.1 Sol, and both are insanely good. Qwen 4 Max Preview went with this warm sepia look and I love it. DeepSeek V4.1 Flash did way better than I expected. Each scene is 1 HTML file, built in DeepSeek Harness on high reasoning. I recorded all 4 myself, moving the light by hand. Music is YuE2, generated locally on my rig.

1:00
Alexey Fateev avatar
Alexey Fateev@superalesha·Nov 16

OpenAI, what the fuck are you guys even doing over there? You give outside experts deep access to your internal models. Cool. Now give your paying users a model that isn't brain dead. It's been forever since your last release that actually felt good to use. Every update since then made it dumber. Astra is dumber than Qwen. Dumber than DeepSeek. Open weights I download for free beat the thing you charge me for every single month. I'm honestly ashamed I still pay for this subscription. The only reason I haven't cancelled is that I can plug it into my own bots and tools. That's it. Not the model, just the integrations. Ship something that works or stop pretending you're the frontier.