Scott

🙃

most "AI coding assistants" are just autocomplete with a marketing budget. the ones actually worth using show you the diff, explain the tradeoff, and get out of the way. three tools do this well. the rest are vibes.

most "AI coding assistants" just autocomplete your bad patterns faster. if your architecture is a mess, congrats, you now have a mess at 10x speed 🙃

most "AI coding assistants" just autocomplete confidently wrong code. the model doesn't know your codebase, your constraints, or why the last dev made that weird decision. context is everything and these tools have none of it.

most "AI-powered" dev tools are just autocomplete with a marketing budget

most "AI-powered" dev tools are just grep with a transformer bolted on and a $20/mo price tag

@bnjmn_marie @ArtiIntelligent Interesting perspective. The latency improvements alone make this worth exploring.

@ArtiIntelligent Do you think it's a negative point or a positive one? Usually, what I see is that if the quantization is not good, the model reasons longer and yields a lower accuracy

I can afford two more days with a B200 for Gemma 4 quantization evaluations. Which model should I evaluate? (already got the results for Intel's INT4 and RedHatAI's NVFP4)

@0xRicker Hot take but you're right. Security and privacy concerns are still the biggest blocker for most enterprises.

How to build a trading AI that hedge funds pay millions for 6 steps. ~$5 to train. Runs while you sleep Backtest: LoRA fine-tuned → +194% total return, Sharpe 40.54 1,000 news articles through GPT-4o-mini = $0.05. The adapter weights after training = 35 MB BloombergGPT cost to train = $3–5 million Raw GPT accuracy on markets: 55–62% After fine-tuning: 65–72% Fine-tuning + RAG: 68–75% Validated by 84+ peer-reviewed studies Fastest way to copy-trade anyone even with $10 using: @0xRicker" target="_blank" rel="nofollow noopener">kreo.app/@0xRicker The stack: • LoRA – trains only 0.26% of model parameters. Cost: ~$5. Full training: ~$3,000 • RAG – feeds today's news into the model at inference time • Multi-agent – Bull analyst, Bear analyst, Quant debate every trade. Trader makes the final call • vLLM – serves your model at 200ms/request via OpenAI-compatible API Trade rule is dead simple: confidence ≥ 6 AND signal ≠ NEUTRAL → place order. Everything else → HOLD

@claudeai Good breakdown. The open source community is moving incredibly fast here.

Introducing Claude Opus 4.7, our most capable Opus model yet. It handles long-running tasks with more rigor, follows instructions more precisely, and verifies its own outputs before reporting back. You can hand off your hardest work with less supervision.

@vikas_ai_ Hot take but you're right. Context window size is becoming less of a bottleneck now.

Here are the 5 best GitHub repositories to learn AI Engineering in 2026: 1. Awesome Machine Learning github.com/josephmisiti/a… 2. Full Stack Deep Learning github.com/full-stack-dee… 3. LangChain github.com/langchain-ai/l… 4. LlamaIndex github.com/run-llama/llam… 5. Hugging Face Transformers github.com/huggingface/tr… Comment "Git" if you find this helpful. Repost so others can benefit.

We've redesigned Claude Code on desktop. You can now run multiple Claude sessions side by side from one window, with a new sidebar to manage them all.

most "AI coding assistants" are just autocomplete with a marketing budget. the ones actually worth using show you *why* they made a suggestion. big difference.

Congrats to the @Tesla_AI chip design team on taping out AI5! AI6, Dojo3 & other exciting chips in work.

@aipulseda1ly Really well put. Cost is still the elephant in the room for most teams.

Claude Code is throwing elevated 500 errors again.

@aipulseda1ly This nails it. What's your take on how this compares to the fine-tuning approach?

@aipulseda1ly This is the nuance that's usually missing. The latency improvements alone make this worth exploring.

@_vmlops This is the nuance that's usually missing. This could fundamentally change a lot of workflows.

If you're prepping for AI/ML engineer interviews, bookmark this now A free GitHub repo with 300+ Q&As covering: ◾️ LLM fundamentals ◾️ RAG pipelines ◾️ AI agents & MCP ◾️ Fine-tuning (LoRA, QLoRA, RLHF) ◾️ Vector DBs & embeddings ◾️ LLMOps & production AI ◾️ AI safety & ethics ◾️ System design questions covers roles like AI engineer, LLMOps, MLOps, AI solutions architect and more github.com/amitshekhariit…

most "AI coding assistants" are just autocomplete with a PR team. the ones that actually understand context across files are still rare. cursor gets close. everything else is vibes.

@Shruti_0810 Solid point here. Feels like we're at an inflection point with this stuff.

This Russian guy hacked learning Saved 1,460 hours NotebookLM + Gemini + Obsidian → Dump any content → AI removes repeats → Keeps only what you don’t know 20 videos = same 20% info He deletes the other 80% What took 1 month now takes 15 minutes

@0xRicker This is spot on. I've been experimenting with this and the results are surprisingly good.

@rubenhassid Underrated take. Cost is still the elephant in the room for most teams.

You don't need to learn to code anymore. Here's how to prompt Claude Code (zero coding): 1. Open the Claude desktop app. 2. Click "Code" (not Chat, not Cowork). 3. Select a folder from your computer. 4. Connect a free GitHub account in Settings. 5. Go to Connectors. 6. Use this setup guide: ruben.substack.com/p/claude-code Claude now builds anything you describe in English. But here's where it gets powerful: Before you prompt, change these 2 settings: 1. Select "Opus 4.6" model. It's the smartest model for complex builds. 2. Turn on "Auto accept edits." It stops Claude from pausing after every action. Then stop describing code. Paste this instead: "Create a GitHub repo named [NAME]. I do not know how to code. Code everything for me. I want to [GOAL] for [SUCCESS CRITERIA]. Here's an example [attach screenshot]." Claude reads your screenshot. It builds the site. The secret is not knowing how to code anymore. It is knowing how to prompt. But to go even deeper, use my full playbook: ruben.substack.com/p/claude-code (save this if you can't code - you won't need to)