Knowix (@knowix_eth)

Pinned Post

Knowix

@knowix_eth

Jul 1

Day 0 of SCRAP TO SPARK. Everyone calls the BC-250 ewaste. AMD silicon ripped out of dead mining server. No support, no manual, no respect. I paid $130 for mine. They're already going for $230. Here's what I'm going to do with it: I'm going to make it buy a supercomputer. The BC-250 will run local models inside a Hermes agent, and that agent will build to earn real money until it can buy the machine that makes it obsolete: a $4,699 NVIDIA DGX Spark. The scrap funds its own replacement. Rules: - no top ups - No side income. If the board can't earn its way to a Spark, it doesn't get one. But first I have to bring it back to life. I've got a used NZXT H1 case for $50 came with a 600W SFX PSU and a dead 140mm AIO. Nothing in it was built to hold a BC-250. So before this thing earns a cent: 1. Revive the dead AIO to cool the board. 2. Mod the case to fit hardware it was never designed for. 2. Hand make custom PSU cables so the H1 can power it without letting the magic smoke out. The jank isn't the obstacle. The jank is the point. This doesn't start with a clean boot and a dashboard it starts with a soldering iron and a case that fights back. I'm documenting all of it. The AIO revival, the pinouts, the first boot, the first dollar. If it works, a board the internet threw away will have paid for NVIDIA's desktop supercomputer. If it doesn't, you watch me fail in public. Either way you get real entertainment. Parts are on the bench. The BC-250 is on its way. This is Day 0. See you tomorrow.

738

Knowix

@knowix_eth

23h

claude code subagents are starting to feel like tiny junior dev teams >i give one goal >one agent delegates >that agent delegates again then i come back to 3 context bubbles, working code, and logs that explain nothing everything runs no idea why half of it exists pre-slack management simulator

175

Knowix

@knowix_eth

Jul 4

the worst part of coding agents isnt the bugs its that you live inside one for months and then cant leave >your shortcuts, >your muscle memory, >your whole workflow belongs to them switching to the better tool means starting from zero again

142

Knowix

@knowix_eth

Jul 3

everyone keeps trying to make one model good at every job. I think that is the wrong future. Leanstral 1.5 is Mistral going the other way: - one open model for Lean 4 proofs. - 119B total, - 6B active, - 587/672 on PutnamBench, - 5 unknown OSS bugs found. most people still underestimate specialist models.

Mert Ünsal

@mertunsal2020

Jul 3

Today, we are releasing Le Chaton L∃∀N, aka Leanstral 1.5. It achieves SOTA performance on graduate algebra benchmarks FATE-H and FATE-X and improves Pareto Frontier on PutnamBench, solving 587/672 problems with a x10 cheaper budget. 🧵

263

Knowix

@knowix_eth

Jul 3

everyone wants a new GPU. meanwhile a 2018 Tesla T4 just ran Gemma 4 26B at 250k context and ~9 tok/s. the useful part is not the benchmark. it is the command, model, and workflow other 16GB GPU people can copy. link in comments👇

305

Knowix

@knowix_eth

Jul 3

Model: huggingface.co/unsloth/gemma… Source benchmark: x.com/analogalok/status/2072…

unsloth/gemma-4-26B-A4B-it-qat-GGUF · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

Alok

@analogalok

Jul 2

Local AI optimization is officially outpacing hardware decay. I spent the last 3 hours building llama.cpp from scratch and benchmarking Google DeepMind’s new Gemma 4 26B A4B MoE on a prehistoric 8 year old, $500 NVIDIA Tesla T4 GPU. The results absolutely break the conventional rules of inference. Here is the raw data running unsloth/gemma-4-26B-A4B-it-qat-GGUF via llama.cpp on a completely free Google Colab Linux (Ubuntu) instance: - 35k context: [ Prompt: 788.8 t/s | Gen: 47.4 t/s ] (-ngl 99) - 80k context: [prefill: 490.8 t/s | decode: 20.4 t/s ] - 180k context: [ Prompt: 242.4 t/s | Gen: 11.1 t/s ] - 250k context: [ Prompt: 220.2 t/s | Gen: 8.9 t/s ] llama.cpp flags: ./build/bin/llama-cli -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf -p "Explain the concept of open source software to a 10 year old." -n 24000 -ngl 99 -c 250000 Yes, I shoved a quarter million tokens of context into a 2018 data center card, and it generated at nearly 9 tokens/sec decode throughput without a single Out Of Memory (OOM) crash. For the hardware nerds, here is the exact environment from my nvidia-smi: - Driver Version: 580.82.07 - CUDA Version: 13.0 - GPU: Tesla T4 (Turing Architecture) - VRAM: 15360MiB (16GB GDDR6) - Power: Sipping just 16W at idle, capped at 70W TDP Google DeepMind really cooked with the Gemma 4 26B MoE (Mixture of Experts) architecture. But the real heroes here are the open source chads. Combining Unsloth's QAT (Quantization Aware Training) quant with the brutal C efficiency of llama.cpp allows us to push 50 tokens/sec on hardware that belongs in a museum. What does this mean for you? 16GB VRAM is the ultimate sweet spot for local AI enthusiasts right now. If you own a single RTX 4060 Ti 16GB, RTX 4070 Ti Super, RTX 4080, the new RTX 5070 Ti 16GB, an older 30 series like the RTX 3080 Ti Laptop card, or cloud GPUs like the A10G and L4 you are sitting on an AI goldmine. You already own the future. You don't need to rent cloud APIs. You just need to compile your tools correctly and let your VRAM do the heavy lifting. If you have a 16 GB VRAM GPU, run this and share your numbers for the community. Model's huggingface link in the comments.

0:10

211

Knowix

@knowix_eth

Jul 2

>this is the broken part of agent memory >not forgetting a fact >repeating the same task 30 times >still not getting sharper >Epoch tested it with a board game >better models started higher >the reps still did not compound

Epoch AI

@EpochAIResearch

Jul 2

Introducing EBR-bench, our new benchmark to measure on-the-fly learning. AI repeatedly plays a challenging board game called Earthborne Rangers and tries to learn from its mistakes. So far: no signs of improvement.

317

Knowix

@knowix_eth

Jul 2

local models have been "good enough" for a while now the problem was never capability. it was that 8 tokens per second makes agent loops unusable ollama just shipped multi token prediction for gemma 4 on mac and the speed difference is not incremental at some point the gap between local and API stops being about quality and starts being about latency and cost i think we just crossed that line for a lot of workflows

ollama

@ollama

Jul 1

Gemma 4 is now nearly 90% faster on Apple Silicon with Ollama using MLX! The speedup comes from improved multi-token prediction (MTP), now on by default for Gemma 4, with more models to come. Ollama automatically tunes how many tokens to draft as it runs, so it never slows generation down when speculation no longer contributes to a speedup.

309

Knowix

@knowix_eth

Jul 1

Tomorrow my token usage will be maxed out

Anthropic

@AnthropicAI

Jul 1

Claude Fable 5 will be available again globally tomorrow. After a series of productive conversations with the US government, we're redeploying the model with a new set of classifiers to target and block more cybersecurity tasks. In the near term, some routine tasks like coding and debugging will fall back to Opus 4.8. We’ll continue to refine these classifiers over the coming weeks to reduce false positives and better distinguish genuine misuse from legitimate requests. We’ve also begun drafting a consensus framework—with Amazon, Microsoft, Google, and other Glasswing partners—for assessing the severity of AI jailbreaks and how AI developers should respond to them. We invite other industry partners and model providers to join us in this effort. Finally, we’re scaling up our collaboration with the US government on model testing and safeguards. This will include pre-release access to models and safeguards for evaluation, information sharing on jailbreaks and misuse, and dedicated resources for joint research. Thank you to our users for your patience, and to our partners across the government, industry, and the research community who worked alongside us to make Fable 5 available again. Read our full blog: anthropic.com/news/redeployi…

376

Knowix

@knowix_eth

Jun 30

Most coding agents are rewarded for producing a correct final patch. Ornith rewards the agent for structuring the work correctly before code starts. The missing training layer is not syntax or implementation. It is task decomposition, ambiguity handling, and execution control.

183

Knowix

@knowix_eth

Jun 30

x just shipped hosted mcp so your agents can read the live feed natively sounds free until you realize every agent call burns paid api credits so now your cursor agent is racking up a tab every time it checks what broke today we went from "scrape and pray" to "metered firehose" in one announcement

Developers

@XDevelopers

Jun 30

Announcing the hosted X MCP. Agents now have access to the best real-time information source in the world. Connect Grok, Cursor, or any MCP-compatible AI tool to the X API without any setup! Check it out here: docs.x.com/tools/mcp

225

Knowix

@knowix_eth

Jun 29

My BC-250 shipped last week. paid $130. now they’re $250 X is treating it like a cheap local AI PC. the real steps/workflow are cursed: >mod PSU >print 3D mounts/case >server board cooling >Ubuntu customization >ECU unlock >model inference tuning. not average user hardware.

186

Knowix

@knowix_eth

Jun 26

This is why Hermes's latest update is interesting. not just because it can beat stronger models on a benchmark. because weaker models can be combined into something more useful. if that makes agent runs cheaper without losing much quality, that changes the whole cost curve.

Nous Research

@NousResearch

Jun 26

The strongest models are gated and access is granted only to a select few. Hermes Agent now exposes MoA presets as virtual models, giving you capabilities beyond the publicly available frontier: 8% higher than Opus 4.8 and 11% higher than GPT 5.5 on our upcoming benchmark.

1:06

198

Knowix

@knowix_eth

Jun 26

GPT 5.6 before GTA 6

OpenAI

@OpenAI

Jun 26

Introducing a limited preview of GPT-5.6 Sol, our next generation frontier model, as well as GPT-5.6 Terra, a balanced model for efficient, everyday work, and GPT-5.6 Luna, a fast and affordable model for high-volume work. openai.com/index/previewing-…

151

Knowix reposted

OpenAI

@OpenAI

Jun 26

Previewing GPT-5.6 Sol: a next-generation model

OpenAI previews GPT-5.6 Sol, a next-generation model with stronger capabilities in coding, science, and cybersecurity, paired with its most advanced safety stack.

openai.com

3,564

5,724

40,349

17,846,242

Knowix

@knowix_eth

Jun 26

there is a new agent benchmark where six LLMs run competing coffee companies for 90 days some models negotiate hard stack inventory and compound profit others spend the whole simulation analyzing their position then go bankrupt doing nothing turns out the failure mode for most agents isnt wrong answers its just sitting there thinking forever

Sakana AI

@SakanaAILabs

Jun 26

SakanaAIは、有限責任あずさ監査法人と共同で、LLMエージェントの長期的な経営能力を評価する新しいベンチマーク「CoffeeBench」を公開しました。ブログ：sakana.ai/coffee-bench/ 現実の経済では、消費者へ直接売るビジネスだけでなく、企業同士が継続的に取引するビジネスも重要です。CoffeeBench は、農家・焙煎店・小売店の計6社が参加するコーヒー業界のサプライチェーンをシミュレーションし、各社をLLMエージェントが運営。90日間にわたって価格交渉・発注・在庫管理などを行い、純利益の最大化を目指します。最新のLLMを同じ環境で競わせると、経営成績は大きく分かれました。積極的に交渉し、利益に直結する一手を打ち続けるモデルがいる一方で、自身の状況を分析しながらも行動に移さず、待機し続けて赤字に陥るモデルも出てくるなど、長期タスクならではの振る舞いの違いが観察できました。 CoffeeBenchは、長期にわたり相互作用するLLMエージェントの能力や振る舞いを評価・分析していくための第一歩です。今後は、複数エージェント間で生じる協調・競争・逸脱行動や、その監査・ガバナンス手法の研究へと発展させていくことを目指します。本研究は ICML2026 Workshop "Failure Modes in Agentic AI" にて発表予定です。論文：arxiv.org/abs/2606.16613 ☕

671

Knowix

@knowix_eth

Jun 26

claude code can act like a production studio not just generate clips research, script, voice, assets, captions, render this is the part of agents people still underrate the wrapper around the model is becoming the product repo in comments

116

Knowix

@knowix_eth

Jun 26

github.com/calesthio/OpenMon…

GitHub - calesthio/OpenMontage: World's first open-source, agentic video production system. 12...

World's first open-source, agentic video production system. 12 pipelines, 52 tools, 500 agent skills. Turn your AI coding assistant into a full video production studio. - calesthio/OpenMontage

github.com

Knowix

@knowix_eth

Jun 25

>stop writing better prompts for Claude Code >start writing better stop conditions >run until tests are green >iterate until lint hits 0 warnings >realize the prompt was never the product >the loop condition was

144

Knowix

@knowix_eth

Jun 24

Everyone on AI Twitter suddenly discovered the BC250 and is acting like it's a $140 RTX 3090. Reality check: >It's a mining APU, not a normal GPU >You need custom Linux setups >Vulkan backend, not the standard NVIDIA stack >Different board revisions exist >Power delivery is weird enough that people literally warn you not to plug the wrong connector into it >Most of the impressive benchmarks are running community patches, custom Mesa builds, GPU unlocks, and a lot of tinkering >Custom power supply wiring needed >3D printed case needed >Need to bypass the factory CPU lock None of this makes it bad. In fact, that's exactly why it's interesting. Getting 16GB of GDDR6 and a fully functional AI node for ~$140 is kind of insane. Just don't expect to plug it in, install Ollama, and magically have a 70B inference server by dinner. But once it`s set up, in theory it can run QWEN 3.6 35B-A3B at around 70 to 100 token/s. I bought one to see how much of the hype is real. Full teardown, setup process, benchmarks, power consumption, and local AI testing video coming soon

271