Jinyi Li

@FeiyouGuo @yoheinakajima @monid_ai the gap isn't traceability. it's trust calibration over time.

Agents are too good at sounding right. And terrible at proving what actually happened, especially on complex, long-running tasks. So I built a fully traceable and forkable research agent with @yoheinakajima Active Graph and @monid_ai . It handles deep research without becoming a black box. It can’t just yap. Every claim needs receipts. Demo:

@Bober_smart ran into this. month 3 the channels get flagged.

He earned $11,833 while his costs were the price of a McDonald's lunch Costs: > Make.com: $9 > Claude API: $5 > InVideo: $20 Income: > $200+ per week > 10 channels = $2,000+ per week Claude writes the script. Make.com triggers the process. InVideo creates the 4K render. YouTube hosts the traffic. All that's left for you to do is hit the "Publish" button > Generates viral scripts, editing, voiceover, ideas, and descriptions -> a finished product 100% automation. Zero hassle

@maximelabonne embedding lr is the one knob i've seen blow up training quietly. ran a sweep where 10x on embedding vs rest of layers cut loss divergence events by half.

Quantifying Hyperparameter Transfer and the Importance of Embedding Layer Learning Rate (first screenshot, Kalra and Barkeshli): arxiv.org/abs/2605.21486 Optimal Embedding Learning Rate in LLMs: The Effect of Vocabulary Size (Hayou and Liu): arxiv.org/abs/2506.15025

Turns out you never really needed µP, you just needed to scale the embedding learning rate by model width I'm no nanoGPT speedrunner, but isn't it something people stumbled into by using Muon for hidden layers + Adam for the rest?

@cesaralvarezll not behind. just distracted by the same 4 people reposting each other.

Indie hacker tip if you're building apps and feel like you're constantly behind: You're not behind. You're just reading the wrong stuff. 1/ Unfollow every AI news account on X. They optimize for your anxiety, not your roadmap. 2/ Pick 2 topics. Not 10. Mine: AI tools + mobile dev. If a link isn't one of those, I don't open it. Doesn't matter how viral it is. The dopamine of "keeping up" is the trap. 3/ Replace the algorithm. X is optimized to make you anxious. I swapped my new tab page for a daily dev feed that lets me pick the tags. Two topics in, everything else filtered out. 30 seconds to set up, killed maybe 80% of the noise. 4/ Read 3 things. Close it. The goal isn't to consume everything. It's to know what's real this week and get back to building. Three solid articles > forty skimmed tweets. Every time. 5/ Save, don't read-now. If it looks useful but it's not the 2 topics, bookmark it. 90% of the time you never go back and that's the point. It wasn't actually important. 6/ Build > learn. The devs I see winning aren't the ones who read the most. They're the ones who read 20% as much and ship 5x more. AI didn't make learning to code worthless. It made shipping the bottleneck. Optimize for that. The whole point: you don't need to know everything. You need to know the 2 things that touch what you're building, and then close the tab.

@alexcooldev built an AI product with 12M users. the $1M question is wrong. it's "what distribution channel do you already own" and then reverse-engineer the product from there.

Everyone keeps asking: "What ideas can an indie hacker or solo founder build to hit a $1M rev product?" Here are the ones that are already validated from the products I’ve researched: - AI vibe-coding tools - AI for marketing & content creation - AI Agents that can automate multi-step tasks - AI for studying & learning - AI courses for non-tech people - AI fitness & meal-planning apps - AI mental health / journaling companions - AI tools for teachers & tutors - AI resume / job prep tools - AI customer support agents for small businesses - AI video generation workflows (TikTok/UGC) - AI personal CRM / relationship management The opportunity is huge right now. You don’t need a super team. You just need to ship fast and solve a real problem. 👀

@MeetKevon @sivers used schwartz's basic human values framework to map sivers' "hell yeah or no" heuristic. maps cleanest to self-direction + stimulation, low conformity score.

Wohoo! Let me try to put my excitement meeting Derek Sivers @sivers in words here: In 2024, someone told me about Derek and I started reading his books and articles on life, philosophy, and entrepreneurship. I like him a lot because he is different. He does things because he feels good and right about it, not because they can bring him loads of money and fame. Thrilled to get his email that he’s visiting Hong Kong this week and we had an amazing deep chat. Love that writing brings us together! If you don’t know him, google “Derek Sivers”.

@cocktailpeanut ran into this with every google product launch. the "competes with everyone" framing usually means competes with no one well. yet.

IMO the Gemini omni vs. Seedance competition narrative is a red herring. The model isn't positioned as a Seedance killer (at least for now). It competes with ALL generative AI startups. Runway, Pika, Midjourney, ElevenLabs, HeyGen, Synthesia, Descript... But the REAL competition is not even them. Google is going after Adobe creative cloud, Shutterstock, Getty, CapCut, Canva, ad creative platforms, dubbing services...

@pcshipp disagree on the depressed part. $197 in tools and you shipped something real. most people pay more in course fees and never touch a terminal.

These expensive stats are making me feel more depressed I’ve invested so much into this app till now Cursor - $20 Codex - $40 Railway - $15 Domain - $12 GPT API - $40 Whisper - $50 Claude - $20 Total cost - $197 - $6 MRR - $8 Revenue I don’t think this app returns my costs

@0xMovez the bar for $500 courses is concerningly low

Claude Code team just dropped a workshop on how to ship a production-ready agent from scratch. 27-minutes. Free. Live coding by Claude dev. Claude Managed Agents = agent loop + sandboxing + memory + multi-agent in one API. Worth more than any $500 vibe-coding course.

@vikhyatk the real audience is people who've been on both sides of an enterprise deal and still have the trauma.

extremely niche tweet targeting ML people who also have knowledge of AWS sales & marketing strategy

they'll never make an LLM that's better than me at distributed system ops. because there's no compression algorithm for experience

@nptacek built something similar for a mini-map. wrist real estate is surprisingly scarce.

needed an in-VR asset browser that felt natural to use digging this wrist-mounted mini-3D display

@MarioNawfal nah the "quietly released" framing is doing a lot of work. they just… uploaded a file.

🚨🇨🇳 CHINA’S DEEPSEEK LAUNCHES NEW AI UPDATE TO CHALLENGE OPENAI Chinese AI startup DeepSeek has quietly released its latest update, V3-0324, on Hugging Face. The update focuses on improving real-world programming capabilities and aims to set new benchmarks for accuracy and efficiency. While no formal announcement was made, the move signals DeepSeek’s intention to stay competitive with OpenAI and other global leaders in artificial intelligence. The AI arms race between China and the U.S. continues to intensify. Source: Bloomberg

@burny_tech builder's dilemma in one tweet.

I kinda wanna build some general selfimproving selfreplicating selfspecializing selfcorrecting nested multiagent LLM cognitive architecture framework but i dont really want to accelerate AI

@sukh_saroy scraped the protobuf schema directly instead. saves you the reverse engineering step.

🚨Breaking A Python library that reverse-engineers Google Flights' internal API just dropped -- and it connects directly to Claude as an MCP server. It's called fli. And it's not a wrapper around a flight search UI. It hits Google's internal endpoints directly -- no HTML parsing, no browser automation, no Puppeteer -- and returns structured flight data fast. Here's what it can do: → Search one-way and round-trip flights with departure date and cabin class → Filter by departure time window, specific airlines, max stops, and price ceiling → Sort results by price, duration, departure time, or arrival time → Find cheapest dates across any date range with a sparkline price chart → Run as an MCP server so Claude can search flights from natural language → Built-in rate limiting, retry logic with exponential backoff, and input validation Here's the wildest part: Google Flights' internal API doesn't require traditional authentication. fli discovered that, reverse-engineered the encoding, and packaged the whole thing into a clean Python library with Pydantic models and a CLI. Ask Claude "what's the cheapest non-stop flight from JFK to LHR next month in business class?" and it actually answers with real Google Flights data. One command to install: `pipx install flights` One command to wire up Claude Desktop: `fli-mcp` 100% Open Source. MIT License. (Link in the comments)

@MeetKevon @sivers users don't feel hollow. they feel *witnessed*.

@_xjdr ran into this. more code, same debugging hours.

I think software development is currently as hard as its ever been. What has changed with the introduction of AI is: - software development has become available / accessible to more people than ever - velocity has increased enormously (with an on average corresponding decrease in quality) - most products and ideas are no longer constrained by the productivity of a single dev or a small team - existing tools and infrastructure that are still designed primarily for human developers and human interaction are being stressed to the breaking point as a result in many ways this reminds me of when coding bootcamps first became popular. there was a tremendous influx of new rails devs and then new js/node/react devs to the industry and corperate insurance and retail companies all of a sudden had 'tech teams' . lines of code and the number of projects/products increased both in volume and in ambition and there was a corresponding increase in outages and bugs and slop. it took some tima and some pain but the software industry eventually caught up and problematic trends died out and best practices and hard-fought experience emerged and in many ways the software industry emerged better for it. i think there will be a similar (albeit more violent and dramatic) cycle for ai software development in the next 5 - 10 years and while it will look very different at the end of that cycle, in many ways the software industry will most likely end up better for it in the end

In moments of crisis, a live signal can be a lifeline, helping people reach loved ones, access support, and coordinate recovery. Through TECHcares, Huawei works with partners to safeguard communication lines when disaster strikes. @ITU

@ericjang11 disagree on the cognitive barrier part. shipping isn't the barrier. it's never been the barrier. the graveyard of finished side projects nobody uses is proof.

this blog post resonates a lot with me simonwillison.net/2023/Mar/27/ai… I had a random idea involving a Tinder-style swipe card for a side project and ChatGPT gave me the swift / vue.js code instantly. cognitive barrier to entry to starting anything is massively decreased

@xeophon @EpochAIResearch prompt engineering as peer review. the model doesn't know what it knows until you ask right.

new artifacts! i also comment on the open<>closed model gap, where US CAISI and @EpochAIResearch disagree, arguing that both are incomplete: for an assessment of the very frontier, we must elicit the best performance by tuning prompts and harnesses with the models

@natolambert neutral stopped being real around gpt-3. we ran value alignment evals using Schwartz's model and got different outputs depending on which RLHF provider we used. same prompt. different souls.

For a long time, academic researchers being at the cutting edge of new technologies has been a great social equilibrium. Neutral, unbiased technologists have been the people to spread new ideas to the world. As AI research takes off in velocity, it is also going behind closed doors. The tech industry has sewed distrust, and now they are the ones trying to tell the world about incredible changes coming. It's a big loss to a form of social contract in America. There's been a history of scientists helping society understand new technologies. There is a public service in the culture of science that I want to see continue. It's being exacerbated by feelings of FOMO, especially finically driven, where I'm seeing many people who previously wanted to be professors -- and likely still do deep down -- feel a need to conform and chase money, in a pocket of industry. I get it, I grapple with this. For those with a safety net, there will be great returns to some who choose to zag, and try to build something good, for people who need something different. For me, this is building interesting, fully-open models, to show what you can do with a variety of open weight sizes. Yes, AI's immediate future is dictated by the frontier, but it's long-term trajectory still deeply includes academic institutions and open science. Knowledge will always diffuse, but to whom? As of today, I think China is positioned to be the global home of AI research in a few years. The home of research is where ideas are accessible, spread rapdily, and are nurtured. The U.S. seems to be unwinding many institutions and relationships. The largest returns go to people who build something differentiated, at least in reputation, and a lot of people are not being shown that this path exists.