Dean Sheppard

One of the weakest models for coding tasks: terrible tool discipline and reasoning you can’t turn off. It’s the only model I’ve managed to push into a looping hallucination just by asking a technical question in the WebUI, repeated three times with a full model reload between prompts. Impressively strong (and wrong) behavior.

GPT-OSS-20B deep dive: 8 coding tasks, 8 passes, 1.8 GB VRAM. >setup: RTX 4060 Ti 8GB, WSL2, llama-server with ncmoe=30, Pi coding agent. model uses 1.8 GB VRAM, 10 GB host RAM. >result: 8/8 pass. every task produced working, tested code. combined with the original benchmark (portscout + logpulse), that is 10/10 completed agentic tasks on consumer hardware. (find all prompts in HF). >what I found: context efficiency: 6-48% of the 32K window used per task. no task exhausted context. the hardest prompt (multi-module with topological sort) used 47.6%. self-correction works: the model found and fixed its own bugs 7 times across 8 tasks. topological sort direction was backwards, fixed. printf format string missing %s, fixed. sed quoting wrong, fixed. no hallucinated APIs: prompt 5 was a trap, "use only the standard library." model used real modules (http.client, json, time, urllib.parse). no fake convenience wrappers. >weaknesses found: edit tool struggles: exact string matching for the edit tool fails repeatedly. model needs 3-4 attempts before falling back to full file rewrites. this is partly an agent-side problem (Pi's edit tool requires exact match), not purely a model problem. directory scan waste: runs "ls -R ~" from home directory, dumping 11K+ lines into context. happened on 2 of 8 prompts. wastes context budget for zero value. rumination: the model debates obvious things internally. "should I use 127.0.0.1 or localhost?" "is name == 'main' correct?" burns tokens but never causes failures. >why it works: 21B MoE with only 3.6B active params. native MXFP4 FFN weights hold quality under Q4_K_M quant. OpenAI's instruction tuning handles tool call JSON cleanly. 1.8 GB VRAM with expert offload leaves massive headroom on an 8GB card. >recipe: - model: gpt-oss-20b Q4_K_M (11 GB on disk) - agent: Pi coding agent (read/write/edit/bash tools) - server: llama-server --jinja -ncmoe 30 -c 32768 -n 8192 - hardware: any 8GB GPU + 16GB system RAM

8x A100, multi instance GPUs, cost around $20k each for 80GB version. Each A100 can be partitioned into 7 instances, so up to 56 instances max from this only server. Each instance rental is from $2.50 to $3.50 per hour, let’s say $3 *56 =$168 per hour or about $4k per day. $120k per month. Electricity and all, it’s still very lucrative business.

> 8 GPUs in one server rig > dude went homeless to build it > electrical bill costs more than rent now > while everyone else pays $400/month to openai > a 2 GPU desktop kills the api bill forever > rtx 4080 super + rtx 5060 ti = 32gb vram > runs qwen 3.6 with 100k context locally > no rate limits, no api keys, no data leaving the room > agents loop 400 times for free > claude opus still wins on hard reasoning > but local handles 90% of daily work > $1,200 setup pays itself off in 4 months > bookmark this and read the article below

Can't trust anything anymore

@RWPhysics Sure, and this “simplification” has nothing to do with a 50 MB bare Python runtime just to execute this printf😂

C++ vs Python 👌🏻

Automotive engineer here. Despite appearances, any part that involves motors is pretty complicated because of the required safety guardrails. It doesn’t need an advanced MCU, but you guys (customers) came with lawsuits about your kids and pets getting injured by closing windows, so we had to add motor current monitoring and the associated anti-pinch logic.

no damn reason a window switch needs this much computer, and thats why its already broken. all the switchgear in my 40 year old acura is OE and WORKS.

@leopardracer Comparing Qwen3.6-27B to frontier models is like comparing apples and oranges. I’m running a local Qwen 35B-A3B as an experiment, but 95% of the real coding work is done with ChatGPT and the $200/month plan.

THIS DEVELOPER HASN’T PAID AN API BILL IN 3 MONTHS. HIS AGENTS RAN 10,000 TIMES FOR FREE he built a local AI lab under his desk two GPUs 32GB VRAM zero rate limits his agents loop 400 times if they want to his coworkers are still watching the usage dashboard the only thing separating them: llama.cpp + llama-swap every prompt stays on his machine every experiment costs $0 zero invoices bookmark & like this before your next API bill hits

It’s not recommended for two reasons. First, it’s a nightmare to rework because you have to apply a lot of heat to compensate for the solid copper heat sinking. Second, without thermal reliefs, your pad shape is defined only by the solder mask. Any solder mask defect around the pads will result in a failed board, since there’s nothing to prevent the melting solder from flowing away from the pad. There are also some annoying defects, like tombstoning caused by irregular heat distribution across copper pours, but your board is probably too small to experience those. Removing thermal reliefs won’t make the ground connection better - it will just create unnecessary complications. Basically, every component should use the default thermal reliefs, and then, only when necessary, you apply an exception rule and use a solid pad connection instead of inherited one.

@SheppaDean Fair point but since the board is relatively small I rather have good ground connections.

Layout done, managed to cram everything onto a 25x30mm board. Heat isolation of the SHT40 is important on a small board like this, no copper pour + cutout slot so ESP32-C6's heat doesn't skew the readings. Next the fun part, 3D renders and case design!

@DabsMalone No, it won’t work. There are multiple issues, but the major one is that once loaded, there’s nothing to keep the motors horizontal to produce lift. With a loaded net, they will flip 90 degrees, and the whole setup will fall like a rock.

Do you think this could work?🧐

@svpino Most of the forced and artificial hype around OpenClaw was about selling you online courses on “how to make money with OpenClaw.”

The “buy a Mac mini or you’ll never make it” crew moved on already. The whole “I’m making $10,000/mo with OpenClaw in a Mac Mini” was a grift, and it’s now dead. What’s the latest now? DGX Spark?

@TheWapplehouse How do $15k side steps sound okay to anyone? Am I the only one too poor for that?

Remember when idiots were buying these things

@PR0GRAMMERHUM0R Oh, the famous 4o! You’re using it wrong! Instead of building a romantic relationship and debating existential topics, you’re just asking stupid math questions.

floatingPointArithmetic

@repairfan_com The tip isn’t missing, it’s stuck inside the fuse holder. Disconnect the power before trying to pull it out.

電源が入らないと言う機器のヒューズを確認すると先っぽが無い😳

@OrganicGPT "Let them eat cake” (c) Marie Antoinette

If you wanna run AI models locally, the best option is an RTX 6000 Pro. DON'T get a 5090/4090. And DON'T listen to people who hype the 3090; those cards are beat at this point. Get the RTX with education discount through Nvidia. These used to be $8000, now they're +$9000.

@witcheer I’d say, look into how the PCIe slots are spaced if you decide to add another GPU one day (it will eventually happen). Most motherboards leave no room for GPU airflow when two are installed.

I'm about to make my next computer hardware purchase. If you know your way around computers, please help me make sure this is a wise choice. it's a significant investment, and this will be my first time buying computer hardware. I currently have: >CPU: AMD Ryzen 5 7600X (6 cores, Zen 4, AM5) >Cooler: Arctic Liquid Freezer II 360 A-RGB + Arctic MX-4 paste >Motherboard: ASUS TUF Gaming A620M-PLUS WIFI (mATX, A620 chipset) >RAM: 32GB DDR5-6000 CL36 (2×16GB Corsair Vengeance) >GPU: MSI GeForce RTX 4060 Ti VENTUS 3X OC 8GB >Storage: Kingston KC3000 1TB NVMe >PSU: MSI MAG A650GL (650W, 80+ Gold) I want to buy: >Used RTX 4090 24GB >PSU 850W 80+ Gold (e.g. Corsair RM850x) >RAM 64GB DDR5-6000 (2×32GB)

@eevblog Their whole business is based on hiding buyers and sellers from each other and making money from transactions. Of course, any link outside eBay triggers alarms. They became paranoid about it a while ago.

You aren't allowed to put PDF datasheet links in ebay listings it seems.

@Soaringeagle45 This cannot be real, it’s already collapsing!

Time for Sheetrock! 💪

@fishPointer And this is actual scalper reasoning. They bought RAM at a fixed (lower) price, then sell it at “market” value, expecting to squeeze every dollar out of the shortage. It’s a greedy practice, nothing more.

RAM (Market Price)

@loktar00 Gemma is too chatty - like an Instagram girl - while Qwen is an introverted handyman dude.

I love Qwen 3.6 but the more I test between Qwen and Gemma the more I respect Gemma's ability to not overthink it and just spit an answer out.

@sirius_kpc They had to strike a balance between having no components at all (which looks suspicious) and using something inexpensive like a 0-ohm resistor. In all fairness, 99% of people don’t even know why a 1W 0-ohm resistor (which looks like a 2512 package) shouldn’t be there.

本来PMIC載せるところに素人はわからんやろwって感じで抵抗載ってるの草 流石に舐めてて面白い