Ying Jin

We (w/ Shansan Gong, Ruixiang ZHANG, Huangjie Zheng, Jiatao Gu, Navdeep Jaitly, Lingpeng Kong) released a family of 7B diffusion language models, DiffuCoder, that specializes on code generation, with a focus on understanding and improving masked diffusion models. A core analysis of DiffuCoder is the autoregressiveness (AR-ness) score, a novel metric that quantifies the causal patterns in decoding, revealing how diffusion models break from strict left-to-right generation for more flexible, non-linear code planning. Recent advances in autoregressive (AR) models dominate code generation, but diffusion-based LLMs (dLLMs) like DiffuCoder offer a promising alternative, especially for complex programming tasks. DiffuCoder explores how these models decode differently—showing less global AR-ness in code tasks compared to math—and how temperature affects both token selection and generation order, unlike traditional AR models. We also introduce coupled-GRPO, a post-training RL method with a coupled-sampling scheme, to reduce performance drops during accelerated decoding, boosting parallelism and efficiency. We use a self-improvement pipeline that leverages AR-ness analysis, coupled-GRPO optimization, and evaluation on benchmarks like AceCode-89k to refine decoding strategies. This approach enables DiffuCoder to navigate diverse code generation pathways and enhance performance with modest computational overhead. Looking ahead, we aim to further leverage Reinforcement Learning to steer code generation through these decoding patterns, with the discrete nature of AR-ness scores providing a foundation for search-based strategies—ideal for the sparse rewards of optimizing complex code structures. Check out our full paper and code for a deeper dive! Paper: https://lnkd.in/gVWU3BDJ Code: https://lnkd.in/gmXTZ_6n Models: https://lnkd.in/gTcKCDr9 #MachineLearning #AI #CodeGeneration #DiffusionModels #NLP

220

5 Comments

👉 My Key Takeaways from Chip Huyen's Recent Interview on Lenny's Podcast Chip Huyen is the author of the widely recognized "AI Engineering: Building Applications with Foundation Models". Link to the podcast: https://lnkd.in/dxZ-tFWX 💡 Importance of post-training.  Pre-training gives you raw capabilities (next token prediction on massive data), but post-training is what makes the model actually usable. SFT on high-quality examples + RLHF. Fine-tuning should be your last resort, not first. Most problems can be solved with better prompts, better data, or RAG. 💡 Evals.  You can't improve what you can't measure. Need multiple types: unit tests (does this specific prompt work?), integration tests (does the whole pipeline work?), regression tests (did we break something?), and user feedback loops. The hardest part isn't writing evals; it's maintaining them as your product evolves. 💡AI products.  Reliability and UX matter more than models. Most AI product failures aren't about bad models: they're about reliability (API limits, latency spikes, poor monitoring) and UX (users don't understand how to use it, doesn't fit workflow). Building reliable platforms and talking to users constantly beats chasing SOTA models. Most insights come from watching users, not from benchmarks. 💡How to improve AI-powered apps.  What people think improves apps: staying current on AI news, chasing newest agentic framework, obsessing over vector database choice, constantly evaluating model benchmarks, fine-tuning models. What actually improves apps: talking to users, building reliable platforms, preparing better data, optimizing end-to-end workflows, writing better prompts. Better prompt engineering beats switching models 90% of the time. A well-crafted system prompt, clear instructions, good examples (few-shot), and proper output formatting can transform a mediocre experience into a great one. 💡 Advice for builders.  Start with user problem, not with cool AI technique. Use the simplest solution that works (often that's a good prompt, not a fine-tuned model). Build evals early. Focus on end-to-end experience. Don't fine-tune unless you've exhausted everything else. Don't treat AI as deterministic (it's not, you need to handle variability). Don't ignore data quality (garbage in, garbage out).

1

1 Comment

Podonos just raised $2.4 million in pre-seed funding, and the signal is loud and clear: the voice AI industry finally has someone measuring the sound, not just making the noise. Based in Los Gatos, Podonos is building the infrastructure layer that decides whether voice AI feels real, human, and ready for primetime. The industry is racing toward a projected $47.5 billion by 2034, but growth without standards is chaos. Podonos is making sure the hype actually holds up when you press play. At the center is ✦Soohyun Bae, PhD in Computer Engineering from Georgia Institute of Technology and Y Combinator W22 alumnus. His track record stretches from engineering roles at Google Maps to leading AR mapping at Niantic, plus co-founding Bobidi and TickTock AI. He has seen what happens when technology scales without proper #guardrails. With Podonos, he's betting the future of AI isn't just about how smart the models get, but how believable and trustworthy they sound. Podonos evaluates models across naturalness, similarity, emotion, recognition accuracy, pronunciation, tone, and resilience. Add in #personaconsistency, and suddenly you're not just testing machines, you're testing their ability to pass as human. The twist? They deliver results in under twelve hours, when the legacy approach drags on for two months. In AI time, that's the difference between relevance and irrelevance. Investors caught the beat. Serac Ventures, led by Kevin Moore, took the lead, backed by NAVER D2SF, the venture arm of South Korea's NAVER, and KAIST Venture Investment Holdings. Together, they've now backed Podonos twice, with this $2.4 million round stacking on top of a $750,000 raise in March 2025. Total funding sits at $3.15 million, and every dollar is pointed toward scaling #engineering, refining #analysistools, and launching into the U.S., Europe, and Southeast Asia. This isn't theory. Podonos already serves Resemble AI, Play AI, and Sanas AI, hitting six-digit ARR by November 2024 with just four employees. Their secret weapon is reach: 150,000 evaluators spread across nine languages and thirteen locales, forming a human-in-the-loop platform that blends speed, rigor, and global coverage. That reach lets Podonos deliver evaluations at a scale and pace that no competitor has matched. The strategy is ambitious: expand into high-demand sectors like #healthcare, #finance, #gaming, and #advertising, then move beyond voice into #multimodal model evaluation, spanning video AI and large language models. The mission isn't just to check if AI works, it's to decide if AI works for people. #Startups #StartupFunding #EarlyStage #VentureCapital #PreSeed #AI #AIInfrastructure #Voice #VoiceAI #VoiceTech #Infrastructure #Technology #Innovation #TechEcosystem #StartupEcosystem #Hiring #TechHiring If software engineering peace of mind is what you crave, Vention is your zen.

6

1 Comment

TL;DR: Cascade RL and even Cascade smaller LR SFT for mitigating catastrophic forgetting. Nvidia recently showed that Cascade RL can mitigate catastrophic forgetting. Really excellent and exciting work! (https://lnkd.in/gDwKxG5f) We also observed similar effects in healthcare domains (clinical trial tasks) in our recent paper: Developing Large Language Models for Clinical Research Using One Million Clinical Trials https://lnkd.in/gNCf9M4i. Beyond Cascade RL, we also find that Cascade SFT with a smaller learning rate can significantly suppress catastrophic forgetting, even across highly heterogeneous clinical trial tasks. This is based on the findings in my previous paper below: SFT Doesn't Always Hurt General Capabilities: Revisiting Domain-Specific Fine-Tuning in LLMs https://lnkd.in/gqWJdvmX.

58

2 Comments

I recently came across an interview with Jensen Huang (NVIDIA’s CEO), and it pushed me to reflect on his thought process. He makes big bets early, often before others believe in the idea and then waits patiently for the payoff years later. Back in 2012, when NVIDIA was mainly known for gaming chips, Jensen doubled down on AI. Many doubted the move. But thanks to earlier groundwork like CUDA (built in 2006), NVIDIA was ready when AI demand exploded. That patience and conviction transformed a risky decision into a $2 trillion success story. Key lessons I take away: Look at where the world is going, not just where it is. Start small but with conviction and keep building even if others don’t see it. Think 10 years ahead: what will people absolutely need, and how can we act today? Accept uncertainty because most bets take time, but the right one can change everything. For me, this mindset isn’t just about tech. It’s about product management, leadership, and building for the future even in the face of uncertainty. #AI

5

DeepSeek V4 drops in upcoming days. February 17th. Lunar New Year. I've been tracking the leaks. The technical specs. The benchmark rumors. Here's what has Silicon Valley engineers quietly stressed: Engram conditional memory architecture. Built for coding tasks. Early benchmarks suggest it outperforms Claude and GPT in specific workflows. Cost to build? $5.6 million. OpenAI spent 100X that on GPT-4. Here's what matters: This isn't about model size anymore. It's about memory architecture. V4 reportedly remembers context across sessions without burning tokens. Every coding conversation builds on the last. For developers, that changes everything. No re explaining your codebase every session. No context window limits killing momentum. No starting from scratch daily. DeepSeek already overtook ChatGPT as the #1 app on US App Store last month. V4 could be the moment coding assistants split into "before" and "after." Prediction? If these benchmarks hold, February 17 becomes the day the AI coding wars got real. Google, OpenAI, Anthropic they're all watching. Sources: TechCrunch, The Verge, LinkedIn Tech Reports (Jan 2026) #DeepSeek #AI #SoftwareEngineering #CodingAI #TechNews

27

24 Comments

🚀 🚀 How I helped build a rocket as a software engineer 🚀 🚀 Full Article: https://lnkd.in/gFavf3eY Last year, Shawn Wei Chew and I had the amazing opportunity to serve as software leads for our school’s rocket team, Rocket Project at UCLA. We wrote code for a hybrid propellant rocket that reached an apogee of 19,200 ft, making it the most successful flight in our team’s history. Coming from a traditional SWE background, we didn’t know what to expect, and from our first outing messing with microcontrollers, multi-threading & communication protocols - we’ve learnt tons along the way So we wrote this article hoping that it would be helpful for prospective club applicants, future batches, or at the very least, interesting to the average reader. Enjoy! Acknowledgements - this article (and rocket) would not have been possible without: • The larger Rocket Project organization, who helped build every single other part of the rocket from scratch that was not code — to me, software is tremendously easier than engineering so I don’t know how they do it LOL • Shawn, my fellow co-software lead and one of my best friends, who helped review and proofread this article. • Angela & Yifan, our amazing electronic leads who helped read this article and put out countless of fires when Shawn and I couldn’t • Our two other software leads who helped build this software system with Shawn and I • My wonderful friends who I roped in to build the web app with us • The electronics team, who were all extremely fun to be around and helped us figure out the hardware side of things so I could just write code

63

11 Comments

The Code Review reported on NVIDIA and Jensen Huang's announcement of the Drive AGX/Alpamayo autonomous technology, and what it means for industry players. With a special Fetii mention! "The announcement arrives at a moment when mobility players like Uber, Lyft, Waymo, and newer models such as high-capacity group transportation platform Fetii are defining increasingly distinct approaches to how people move." As NVIDIA and others standardize and open up autonomous vehicle technology, the cost curve will continue to compress and adoption will accelerate across OEMs. Autonomy won’t remain a novelty or a premium feature, it will become table stakes. The real advantage will belong to platforms that can turn that technology into sustained, high-utilization demand. That’s where Fetii is uniquely positioned. By focusing on high-capacity, repeatable group movement, Fetii and its partners are built to extract real economic value from autonomous fleets as they scale.

23

3 Comments