Neo Kim

How to become a 10x software engineer (in 2026). Read these 10 books:

@stats_feed live in a cold country

People who rarely get sick, what are your secrets?

@aibrandscaling @claudeai What are you building

@systemdesignone @claudeai And I'm out of credits 😂

We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.

How to become a 10x software engineer (in 2026). Read these 10 books:

As an AI Engineer. Please learn LLM Evaluation Concepts: >Eval: A structured test to measure the quality of LLM outputs >Criteria: What "good" means for your use case >Rubric: A checklist to turn vague criteria into specific, scorable questions >Test case: One input plus an expected/scored output >Golden set: Your collection of trusted test cases built from real user queries >LLM as judge: A strong model scores outputs against your rubric, so you can evaluate 1000s of cases cheaply >Human evaluation: People manually score outputs. It's slow and expensive, but the closest thing to ground truth >Heuristic checks: Simple code checks such as valid JSON, length limits, required fields >Semantic similarity: Uses embeddings to check if two texts mean the same thing, even with different wording >Pairwise comparison: Show a judge two outputs and ask which is better >Judge calibration: Checking how often your LLM judge agrees with humans before trusting it >RAG triad: Did retrieval find relevant context, is the answer grounded in it, and does it address the question >Offline vs Online evals: Offline runs before deployment on your golden set; online samples live production traffic >Regression testing: Rerun your eval suite on every prompt or model change to catch quality drops >Benchmarks: Public exams such as MMLU (knowledge) and HumanEval (code) for comparing models x.com/systemdesignon…

👋 PS - Want my System Design Playbook for FREE? Click the link below to join my newsletter right now: → newsletter.systemdesign.one/join (200K+ software engineers have already signed up.)

♻ RT to help others become 10x software engineers 👤 Follow @systemdesignone + turn on notifications.

@stats_feed @grok is this true?

2020 was 2020 days ago.

@IAmSteveHarvey one tesla

If you could own any vehicle you wanted, what would it be?

@GoogleAIStudio are people still vibe coding in 2026?

What are you vibe coding this weekend?

@claudeai while(true):

@Kisalay_ LLM evaluations are very crucial in enterprise settings.

LLM evaluation is becoming one of the most important skills for anyone building with AI. Most people just prompt and hope. The ones who actually measure quality, build rubrics, and use LLM judges are the ones who will ship reliable and trustworthy systems at scale. Time to follow me neo, hahah just kidding😊

@swapnakpanda They're probably big enough to handle the fine (considering the profits).

In 2010, Aaron Swartz downloaded 70GB of articles from JSTOR. He faced $1M fine and 35 years in jail. He took his life in 2013. Meta too downloaded over 80TB of books from Anna's and LibGen to train their AI models. But they never received any punishment. Isn't it illegal too?

@bspectacledGOAT @swapnakpanda He's one of the cofounders of Reddit

@swapnakpanda sorry about Aaron! but the system favours the rich!

@e_opore you're very welcome.

@systemdesignone Thanks for sharing,Neo!

@swapnakpanda thanks..

@systemdesignone Top books for sure.

@csaba_kissi 🤝

@systemdesignone Hope the future engineers dont fully rely on AI

@Prathkum ah, time to change that

@systemdesignone it's been a while since I last read a tech book.

@csaba_kissi looking at this is good for eyes

The open-source library of CSS text effects now contains 66 effects (+9 effects added). You can copy CSS to the clipboard or the AI prompt (ready for your AI tool). text-effects.colorion.co

@Parul_Gautam7 we need one for table football too

Most video-action robot models take an off-the-shelf video generator built for content creation, then adapt it with action modeling. LingBot-VA 2.0 does something different; it pretrains the entire stack from scratch, built for control from day one. → A semantic visual-action tokenizer puts world states and actions in one shared latent space → A causal diffusion transformer trained forward in time, not retrofitted from a bidirectional model → Control knowledge learned at web video scale, not limited by scarce robot demonstration data This fixes three real limitations of the adapted approach. Latents built for appearance instead of dynamics. Inference is too slow to close the control loop. And a pretraining objective that never actually teaches how actions change the world. Going native means the model's priors are built for control from the start, not inherited from a video generator and eroded during a retrofit. Robbyant, an embodied AI company under Ant Group, is building one brain for all robots. @robbyant_brain Project page : technology.robbyant.com/lingbot-va-v2 Paper: github.com/Robbyant/lingb… #Robbyant #Lingbot #ad #Ai