Own Your Compute @OwnYourCompute

Link to Dessa: block.xyz/dessa

Holy shit! Fully local duplex speech model that can runs on your Mac! @jack’s Block just dropped the most authentic full-duplex speech model ever built called Dessa. It was trained from scratch on pure human conversation. No text. No pretrained models. It listens and talks at the same time, interrupts naturally, laughs, catches sarcasm, and runs locally on a MacBook Pro M4 Max.

This is what AI voice should sound like. Natural, imperfect, human-like speech instead of the polished and refined announcer voice .

Own Your Compute @OwnYourCompute

4 hours ago

Holy shit! Fully local duplex speech model that can runs on your Mac! @jack’s Block just dropped the most authentic full-duplex speech model ever built called Dessa. It was trained from scratch on pure human conversation. No text. No pretrained models. It listens and talks at

1 0 1 22 1

6× NVIDIA CMP 170HX GPUs running at nearly 250W EACH — ~1,500W total, 98–99% GPU utilization, yet staying at just 61–65°C! 🔥 The secret? Custom 3D-printed cooling shrouds I designed for 9733 blower fans. I now make them for other CMP 170HX owners too! 🛒 Available on my Etsy store — link in bio. #CMP170HX #3DPrinting #LocalAI partypailsto.etsy.com/listing/455659…

keyboard_arrow_left Previous keyboard_arrow_right Next

The CMP 100-210 is a mining card on the same GV100 silicon as a Tesla V100. Used V100 16GB cards now go for ~$300–500. CMPs are usually cheaper, but the real difference is the firmware locks. Stock CMP: • PCIe Gen1 x1 • Tensor + FP64 throttled to ~1/16 speed Unlocks (experimental) have shown ~14–15× gains in those areas on real hardware, getting surprisingly close to proper V100 numbers. Caveats still apply: not every card unlocks the same, risk of bricking, and some stay limited on SM count or PCIe width. A real V100 is the safer, more supported option if the prices are close.

@gospaceport Kind of regret selling my two 3090 two months ago😭

@deedydas @abliteration_ai Stop shilling for your own investment

@LLMJunky What hardware are you getting?

@dom_kwok Other displaced workers will just join these fields and push down the wages.

@aiwithadb For subscription plans, Chinese models are not that much cheaper than frontiers.

What if future robbers also use robot to rob people 😭

@SeanSemanko At its current form, AI is best for anything computer related like coding. I find other areas such as research lacking.

💸 $10,000,000 on AI tokens in one month. That’s what the OpenClaw creator reportedly spent in September alone. Enough to pay an entire office of workers for a month… just burned on API calls. This is why a lot of serious builders are moving as much as possible to local models. Once the workload is heavy enough, the cloud bill stops making sense — especially for coding agents, research, and anything that runs 24/7. You can still use the frontier models when you need them. But the daily grind? That’s increasingly running on your own hardware. Have you hit the point where API costs started feeling painful?

BBC Cybersecurity expert: "I didnt even know PHP existed until yesterday morning."💀

$225 Tesla P40 vs $1,300 RTX 3090 — cheap VRAM is a trap if you ignore the rest of the build ⚠️ The P40 looks unbeatable on paper ($9/GB). Reality is messier. Here’s the actual breakdown from today’s listings: 💰 Budget tier 🟢 Tesla P40 24GB — $225 shipped Passive server card. Needs real airflow + power mods. Confirm software support for your stack before buying. Not a drop-in. 🖥️ Used consumer cards 🟡 MSI RTX 3090 Gaming X Trio — $1,240 local / $1,300 shipped Dead HDMI (DisplayPorts work). Ask for current load tests. 🟡 HP OEM RTX 3090 — $1,300 shipped Condition details incomplete. Demand proof it works. 🟡 ASUS ROG Strix 3090 OC — $1,250 local / $1,350 shipped Missing screws (was watercooled). Verify load test results. 🧠 Bigger memory 🔵 AMD Instinct MI210 64GB — $3,900 shipped Better $/GB than the A6000s, but only if your software actually runs well on ROCm. 🔵 RTX A6000 48GB — $3,500–$4,050 shipped Higher cost per GB. Used for inference in one listing. 💡 My take The $225 P40 is interesting for a deliberate budget build. It is not a substitute for a tested 3090. At ~$1,300, the 3090s need different questions answered — one has a disclosed fault, another has incomplete condition info. Negotiate. Get load tests. Factor in cooling, power, and software fit. $/GB alone is a bad filter. Owners who’ve run both a P40 and a 3090 for local inference: which model/backend made the upgrade worth it, and how much did cooling + setup actually add to the P40? 👇

Unlocking Full BF16 Tensor Speed on RTX 3000 Ampere GPUs — The 170HX Technique Comes Home NVIDIA shipped every GA102/GA104 Ampere card (RTX 3090, 3080, 3070 Ti, etc.) with a quiet software throttle on tensor math that accumulates to FP32. The driver even admits it: FMLA32 = 1/2. Pure BF16 tensor pipes were stuck around half the silicon’s real capability. On a 3090 that meant roughly 82 TFLOPS instead of ~155. Same cards, same power limit, same driver — just half the rate on the path most modern training and mixed-precision inference actually uses. The fix arrived via a technique refined on the CMP 170HX (the GA100 mining cards that the community has been unlocking for full SM throughput, memory geometry, and more). Those cards use the same style of speed-selector registers (SS0/SS1) and Platform Lock Manager controls. Zagitta’s gpu-unlock Linux kernel module adapts the V67/PLM approach so it runs before the stock NVIDIA driver binds the card, writes the full-rate selectors (SS0 = 0x88888888), performs a function-level reset, and then gets out of the way. Early results from people running the module: 👉 BF16 tensor pipe: 82 → 155 TFLOPS (~1.9×) 👉 cuBLAS BF16 GEMM (8192³): solid gains in the 1.5× range 👉 Real inference prefill (e.g. Qwen-style workloads on 3090s): typically +10–12% end-to-end, because GEMMs aren’t the entire pipeline Outputs stay bit-identical to stock runs. It is opt-in (gpu_unlock.targets=all or specific PCI addresses on the kernel command line), requires a proper cold boot in many cases, and currently targets Linux with GA102/GA104 GeForce/CMP cards. The module itself is not a patched NVIDIA driver — the stock one still loads afterward. This is pure community reverse-engineering of fuses and selectors that NVIDIA left in a restricted state on consumer silicon. As with any such unlock: your hardware, your risk, verify logs (journalctl -k | grep gpu_unlock), and expect the usual caveats around initramfs ordering and power-state quirks. The 170HX unlock work just paid dividends on the cards most people actually own. Full-rate BF16 on Ampere 3000 series is real, measurable, and now available without flashing VBIOS. Repo: github.com/Zagitta/gpu-un… Big thanks to the 170HX researchers and everyone who measured and shared results.

Two out of four isn't bad... (sound on!)

@Polymarket Actual Footage of the garbage