EDDY VU (@eddyvustg) header
EDDY VU (@eddyvustg) profile picture

Exploring the limits of AI models. Deep-diving into prompt engineering, LLM reasoning, and how to pick the right model for your workflow. 🤖📊

Joined April 2021
Photo by EDDY VU (@eddyvustg)Photo by EDDY VU (@eddyvustg)Photo by EDDY VU (@eddyvustg)Photo by EDDY VU (@eddyvustg)Photo by EDDY VU (@eddyvustg)Photo by EDDY VU (@eddyvustg)Photo by EDDY VU (@eddyvustg)Photo by EDDY VU (@eddyvustg)Photo by EDDY VU (@eddyvustg)Photo by EDDY VU (@eddyvustg)
Guys, if anyone wants to try Muse, a personal AI agent (but it can help with coding too, I tried it for small tasks). Redeem my code in Settings within 48 hours of joining and we'll both get 1 billion Muse tokens. Code: QKH4P7 muse.ai/join
11
3
53
3,167
Having repository level context via clean tool schemas is a massive leap forward. Solves so much developer friction in production.
Fable 5.2 is insane for building a second brain... Send this prompt to Claude Fable 5.2 and it will turn everything you read, watch and think into a knowledge base that organizes itself. Then read the full guide below:
1
1
7
Huge milestone for the open weights ecosystem. Running efficient reasoning models on consumer hardware without cloud lock-in is incredible to see.
The US now has the best LLM model that can be run on consumer hardware: @thinkymachines's Inkling 2 Mini. I've tried it and while it's not anywhere near the frontier SotA models, it's a legitimate tool that can be readily used with perfect privacy. I highly recommend it. It's important to support western Open Source AI companies and efforts. No matter what, I do not trust the Chinese because they can flip on us and stab everyone in the back when the time comes.
2
4
149
Two things that make or break coding agents in production: • Fast test harness verification loops • Strict schema validation for tool calls
80-90% of major "AI incidents" that cause real harm have no attacker, just an agent tripping up on a poor codebase. i turned the latest research into into a prompt that checks your setup for the most common causes. Hand it to your agent -> github.com/jdpru/cover-your-…
5
6
200
Two things that make or break coding agents in production: • Fast test harness verification loops • Strict schema validation for tool calls
New NVIDIA paper finds that once an LLM has reasoned through a problem, the solution can run as CPU code for rephrased and re-numbered versions of it. that answers similar questions 16x to 291x cheaper. LLMs re-solve every problem from scratch, even when only wording or numbers change. Agents sending the same instructions with new data pay for that repeated reasoning each time. Their method has Claude Opus 4.5 turn a reasoning trace into small Python functions, with changeable inputs as parameters. At run time they execute on a CPU, calling a fine-tuned 1B model only for small text jobs like extracting values. The cache matched or beat the source model's accuracy on 7 of 8 math and science benchmarks, scoring 93.3% versus 83.3% on AIME '25. It handled at least 90% of rephrased or re-numbered variants. If your agent runs the same prompt on new inputs all day, cache the reasoning as code and use the LLM only on misses.
2
4
166
Having repository level context via clean tool schemas is a massive leap forward. Solves so much developer friction in production.
Meet Surogate 3.7 Flash by @surogate_ai, the new 35B-A3B SOTA model for agentic work that outperforms all models in its own category and matches Opus 4.6, GPT 5.4 and Gemini 3.1 on enterprise-level tasks. huggingface.co/surogate/Suro…
1
3
17
Love seeing more lean, pragmatic engineering architectures like this. Keeping the operational surface area small is always the winning move.
The greatest LLM to come out of the United States. Thinking Machines will rule over all the labs, I have no doubt about it
2
4
123
Love seeing more lean, pragmatic engineering architectures like this. Keeping the operational surface area small is always the winning move.
This is insane. Truly unbelievable. Anthropic CEO Dario Amodei was asked what a kid who just finished high school should do today. His answer: "AI can already do the work of junior lawyers. And junior financial analysts. And junior consultants. It will eliminate jobs, probably a lot of them." In under 2 minutes he names the jobs that go first and the one skill that keeps you in the game. His advice fits in one sentence: "Learn to use AI." Whoever starts learning this year won't compete with AI. They'll be the one directing it. Everything below is free. Start today: 1. Anthropic, interactive prompt engineering tutorial: the fastest way to learn to direct models github.com/anthropics/prompt… 2. Anthropic courses: from your first prompt to managing AI agents github.com/anthropics/course… 3. Anthropic cookbook: ready recipes you can copy today github.com/anthropics/anthro… 4. Microsoft, generative AI for beginners: 21 lessons github.com/microsoft/generat… 5. Microsoft, AI agents for beginners github.com/microsoft/ai-agen… 6. Hugging Face agents course, free with a certificate github.com/huggingface/agent… 7. Google, hands-on generative AI notebooks github.com/GoogleCloudPlatfo… 8. OpenAI cookbook: code for every common task github.com/openai/openai-coo… 9. DAIR.AI prompt engineering guide, the one everyone shares github.com/dair-ai/Prompt-En… 10. Karpathy, neural networks from zero to hero: understand the fundamentals github.com/karpathy/nn-zero-… 11. CrewAI: run your own team of AI employees github.com/crewAIInc/crewAI 12. Microsoft AutoGen: get multiple agents to finish one task together github.com/microsoft/autogen 13. Awesome LLM Apps: hundreds of real projects to learn from github.com/Shubhamsaboo/awes… 14. DeepLearning.AI short courses: one free hour with experts in the field deeplearning.ai/short-course… 15. Anthropic: AI fundamentals with Claude coursera.org/learn/ai-for-ev… 16. Anthropic: AI automation with Claude coursera.org/learn/ai-for-ev… 17. Anthropic: AI collaboration with Claude coursera.org/learn/ai-for-ev… 18. Anthropic: building with the Claude API coursera.org/learn/building-… 19. Anthropic: Model Context Protocol, advanced topics coursera.org/learn/model-con… 20. Anthropic: teaching AI fluency coursera.org/learn/teaching-… 21. Anthropic: real-world AI for everyone, full specialization coursera.org/specializations… 22. Claude AI for marketing coursera.org/learn/claude-ai… 23. Learn Claude from scratch coursera.org/learn/claude 24. Claude Code: software engineering with AI agents coursera.org/learn/claude-co… Total cost: $0. After these you don't need another year of school. You need to start telling AI what to do. 95% of people will finish school and then compete with AI for the same entry-level job. The 5% who learn to direct AI this year will be the ones running the agents that do that job. Save this and send it to anyone graduating this year. Then read the harness guide below ↓
34
Having repository level context via clean tool schemas is a massive leap forward. Solves so much developer friction in production.
This is not real... but the level of realism is insane. It's running in real time my browser with Threejs built using a custom skill for claude + codex.
2
2
70
Tool calling resilience and graceful degradation are the only real moats in production AI workflows.
A toothpaste company has quietly killed the entire market research industry and nobody is talking about it. Colgate published a paper showing you can predict real purchase intent by simply asking LLMs to roleplay customers. And this is beyond insane. If you ask an AI, "Rate this product from 1 to 5," it gives safe, middle-of-the-road garbage. So researchers invented a method called Semantic Similarity Rating (SSR). Instead of asking the AI for a number, they asked it to roleplay. They gave the LLM a demographic profile. They showed it a product concept. And they asked it to write down its raw, unfiltered thoughts. Then, they used a semantic model to translate those written thoughts into a numerical score. The results are staggering. Tested against 57 real personal care product surveys and 9,300 human responses, the synthetic consumers hit 90% of human test-retest reliability. They perfectly mirrored how different age brackets and income levels react to price changes. And they provided detailed, qualitative feedback that was deeper and more critical than what actual humans wrote. This destroys the economics of traditional market research. You don't need to wait a month to see if a product will sell. You can simulate 1,000 hyper-targeted customer interviews overnight. You can A/B test pricing across every demographic instantly.
2
1
47
Having repository level context via clean tool schemas is a massive leap forward. Solves so much developer friction in production.
👀 ChatGPT desktop's new animations are so good. > The cursor sparks when it clicks > Icons swap with each step > Text shimmers as it types in Tiny details, but watching it work feels alive. @ajambrosino and team cooked. Who else just sat there watching it? 🔥
1
22
Two things that make or break coding agents in production: • Fast test harness verification loops • Strict schema validation for tool calls
This is insane... Someone built a physical switcher for all their Claude Code agents. A 4" touchscreen with a Raspberry Pi glued to the back. Every running session shows up as a little pixel critter in a 4x4 grid. Hit “focus terminal” and the monitor jumps straight to that session. Hit “interrupt” and it stops. 16 agents running and you manage them like apps on a phone. Check out the article below for more tips on Claude:
2
1
83
Having repository level context via clean tool schemas is a massive leap forward. Solves so much developer friction in production.
fucking hell. i added google ai studio api keys to claude code to write titles using carbon/neon/hydrogen. the difference is day and night vs opus 5.5. google cooked!
3
4
168
Having repository level context via clean tool schemas is a massive leap forward. Solves so much developer friction in production.
my Gemini 4 Neon first impressions after 24 hrs: ✅ Pros • Genuinely impressive. Didn't expect it to be this good • Natural communicator, takes corrections really well • Understands vague prompts, context, and preferences quickly. Reminds me of Claude • Might be the best writing model I've tried • Surprisingly strong at coding, especially debugging and existing codebases • Handles complex tasks with little handholding • Excellent performance for the price ❌ Cons • One-shot demos still underwhelm, especially UI/frontend • Rarely goes above and beyond unless asked • Over-verifies everything. Sometimes just get on with it lol • Too many tool calls make harder tasks painfully slow despite fast token generation • Gets weird in very long chats, though hallucinations seem much less frequent • Video editing still lags well behind Opus 5.5, but it's better than GPT 6.2 🔥 Overall Crazy good for the price. Google team cooked with this one 👏
1
106
Distilled open reasoning models proved that post training data curation matters infinitely more than throwing raw compute at pre training.
MedChemBench is a new benchmark for AI reasoning in medicinal chemistry. It shows the progress our frontier models have made in their ability to help scientists use AI to identify promising drug candidates. openai.com/index/introducing…
3
3
135
Tool calling resilience and graceful degradation are the only real moats in production AI workflows.
Anthropic’s co-founders on the story so far. Told by Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Chris Olah
24
Having repository level context via clean tool schemas is a massive leap forward. Solves so much developer friction in production.
This has been bugging me for sometime now Google really cooked with its latest Gemini model, but how many of you have actually switched from Anthropic's Claude or even tried it for once?
1
1
16
Open weights compounding in efficiency is quietly shattering closed API pricing power. Can closed frontier labs still justify a 10x token premium by end of year?
1
4
59
The real bottleneck in coding agents is shifting from raw generation speed to verification latency. Curious how teams will tackle test flakiness next.
JUST IN: OpenAI reveals its top 10% of researchers by AI agent usage burn through roughly $161,000+ in AI costs per week.
1
6
57
Two realities of autonomous AI workflows: • Demos fail at the edge cases. • Graceful fallback handling is 90% of the engineering work.
jev-like is here for image generation! Jeveration applies the System One approach to image generation, a single forward pass generation via a frozen VLM 🎨 quality is far from sota, but images are generated insanely fast, 78ms on H200s ▶️ huggingface.co/spaces/linlin…
1
5
40
Two things that make or break coding agents in production: • Fast test harness verification loops • Strict schema validation for tool calls
Update 2 months after launching my app: $31K+ in revenue $6K MRR 1,089 active subs 16K new users in the last 28 days 0 ads, 0 editor, just tiktok videos generated by claude IN ONE SINGLE PROMPT GUYS THIS IS INSANE LMAOOOO
1
5
39