Bijun He

📌 Trends, Judgment, and Leverage: The Long Game in the Tech Industry This year, I was fortunate to be selected for the early-stage startup employee track of Tsinghua University’s Lighthouse Mentorship Program. As part of the program, we hosted a panel at Foothill Ventures last Sunday. I was honored to moderate the conversation with: Chenxi Wang, Ph.D. — General Partner at Rain Capital Michael (Xuecheng) Jin — Founding Partner at Foothill Ventures James Chen — VP of Engineering at Netskope Wenjia Fang — Former Google & Nokia Executive Chuck Huang — Founder at CITCON We discussed what actually compounds across cycles in the tech industry. A few takeaways: 1. What capabilities actually compound over time? Stay curious, keep experimenting, and learn through feedback. Once you understand the underlying dynamics, experience compounds across roles. 2. In tech, how does influence actually scale?  Do great work and speak up when it matters. Move fast, learn from feedback, and stay active in the ecosystem. Build your profile and be ready to capture non-linear opportunities. 3. How can women and minorities break through? Don’t limit yourself. Seek out mentors and people who open doors. Take advantage of the diverse environment and lean into your strengths, such as collaboration, communication, and empathy. These human skills are becoming even more important in the age of AI. 4. In early-stage VC, what signals matter most? Technology risk is often overestimated. Strong leadership, an experienced team, a large market, and the right timing tend to matter more than people think. Thanks to Qi Deng, Jerry Hu, and Chewie L. for co-organizing this event. Great discussion, and many thanks to the mentors and participants for the thoughtful exchange!

43

5 Comments

I wrote an article on how we translate user-generated content at scale at Mercari. It summarizes two years of work with various models, LLMs but not only, and how cost evolved, reducing it 100x over the period. I also go over non-AI features and user experience. https://lnkd.in/gbpQAQmX

39

2 Comments

This is exactly the kind of transformation we help SMEs with at Connective Labs. Different scale, same principle: find the expensive manual process, automate it, keep what makes you you.

9

Honored to have been invited to fal and Henry Shi’s Lean AI Leaderboard Celebration Dinner, celebrating some of the leanest and fastest-growing startups in SF. Here are some of the most overwhelming trends that were discussed among high-growth companies working with small, high-output teams: 1/ Headcount is no longer used as a success metric. As Henry Ward said, “Hiring means we failed to execute and need help.” Headcount is not a flex measure anymore. People aren’t asking “How big is your team?” anymore. They’re asking “How many customers do you have?” 2/ They're finding ways to leverage AI across every function. Every engineer uses Cursor or Windsurf. Every GTM leader is in Clay. The most efficient teams are built AI-first across the board. Thank you to Henry Shi, Burkay Gur, and Gorkem Yurtseven for putting this all together!

67

Many founders and tech professionals are starting to explore the O-1 visa, often assuming that building a strong O-1 profile can be done quickly. However, based on my conversations with Jason Cheung about EB-1A and O-1 profile-building, this is not the case. Even O-1 profile-building can take months or even several years. If you want to start building an O-1 profile, begin by: - Getting involved in activities relevant to the criteria you want to target. - Speaking to reporters if you are already connected to them through your network or company. - Attending conferences to speak, which can help demonstrate your expertise in your field. To be notified first-hand about all our future opportunities like these, subscribe to our calendar of events: https://lu.ma/ventureminer Additionally, if you're interested in learning more about EB-1A and O-1 profile-building, check out Jason's profile-building opportunities database, private group, and other resources through his newsletter (only $12 per month): https://lnkd.in/deqhfpRs #eb1a #o1visa #greencard #swe #tech

13

Finally had some good news on a long-running case with the Amazon Japan Electrical Appliance Approval team. After several rounds of back-and-forth messages, updated test documents, and clarifying a few technical details, we did it. And honestly—it was refreshing to feel that a real person was actively taking care of the case. The product in question is an extension cord, and Amazon required the following: 1. Copy of the METI notification (stamped with the METI receipt seal + model classification table) 2. Photo of the rating plate on the product itself, showing the PSE mark, notifying business name, rated voltage, etc. 3. Conformity Inspection Certificate from a registered inspection body, verifying the model number on the rating plate 4. Self-inspection records for all finished products Our biggest hurdle? We wanted to use the same ASIN as the US listing, so that 60K+ reviews would carry over to Amazon Japan. But… Amazon spotted two issues: 1. The Japan version had a different model number than the US product linked to the ASIN. 2. The US product uses a 3-pin socket design, while Japan requires a 2-pin socket. This led to five separate email exchanges with Amazon. Eventually, they agreed that it is the same product—just localised for Japan’s electrical standards. The last remaining step is approval to sell on our Rising Sun Commerce Amazon Japan seller account. We’ve submitted a purchase invoice to confirm ownership of the products, and once that clears, we're fully over the line. It’s been a hassle, but we’re nearly there—and it’s a big win for the client.

2

📣 PFN has raised an additional 5 billion yen in an extension round (total of 24 billion yen in this latest round). The raised fund will be used for the development of energy efficient MN-Core™ series of AI chips, PLaMo™ LLM, collaborations with the investors, and more🚀 https://lnkd.in/ge4PTzMW https://lnkd.in/gFFnxJGw

33

The relationship between HTX (Home Team Science & Technology Agency) and.Mistral AI began modestly in a small café at NVIDIA GTC, just a year ago. It was there that Chee Wee Ang first met Arthur Mensch . What started as a conversation quickly revealed shared interests, common concerns, and, most importantly, a mutual conviction to build and progress together for Singapore and for France respectively. One year on, that conviction has translated into tangible outcomes. HTX has successfully pre-trained its own Phoenix model to support Singapore’s public safety mission. In parallel, together with Mistral, we co-developed Forge: a testament to what true partnership can achieve. This is not a conventional vendor-client relationship. It is a collaboration grounded in co-creation: developing, debugging, and bringing Forge to market together. In doing so, we have strengthened our intrinsic capabilities and created multiplicative value far beyond what either could achieve alone. Both the journey and the outcomes have been deeply rewarding. We look forward to realizing the many more lighthouses ahead. Geoff Soon Wei Cheng Wilson Tan Wang Jiale Gee Wah N. William El Sayed #HTX #HTxAI #MistralAI #xCloud

55

Joined the AI Accelerate founder session last week at NUS Enterprise together with Murli Ravi from Tin Men Capital and Jeffrey Paine from Golden Gate Ventures to discuss how we nvest in the era of AI. The room was filled with 25 AI-native startups, many of them preparing to raise. A big part of the conversation focused on how investors are currently evaluating AI companies and what really matters beyond the initial excitement. One observation from what we are seeing at 1MX AI is that AI is shortening product development cycles, but it has not shortened the timelines required to build durable companies. Today small teams can prototype and launch AI products far faster than before. Open source models, APIs, and developer tooling mean capabilities that once required large teams can now be built in weeks. But the harder parts of company building remain the same. Customer adoption still compounds slowly, distribution still takes time to build, and long term advantages tend to come from data, workflow integration and user trust rather than model access alone. That is why speed of execution matters, but durability still matters more. Great discussion and thoughtful questions from the founders. Thanks to Jit Singh and the NUS Enterprise team for bringing everyone together! 🚀 1MX AI

57

2 Comments

Insightful reflections on how past transformations sparked by technology breakthroughs became really impactful only when users abandoned their old habits — ie. treating the new tools as merely an upgrade (better, faster, cheaper) of their old ones — and instead thought more deeply about how freedom from old tools’ constraints could empower individuals, companies, societies and cities.

3

Last week, Signifyd’s AI Lab, Data Science, and Engineering teams hosted Dr. Hui Wang for a fireside chat. It is rare to find a leader who can seamlessly pivot from the theoretical "Innovation Paradox" of LLMs to the practical challenges of leading through a period of rapid change. She provided a roadmap that balanced radical transparency with clear executive boundaries, covering an incredible range of ground in just under an hour: * The Productivity Paradox: AI can speed up coding by 50%, yet corporate impact often stays in the single digits. Why? Because the bottleneck often isn’t the tech, it’s the "human maneuvering." Without bridging the gap between decision-makers and data, the code can’t save us. * The "Default" Trap: When data is limited, the "truth" is often hidden. Without intentionality, we will automate the invisibility of those who don't fit the mold. That is where human intuition becomes a technical requirement. * Sponsorship > Mentorship: True growth doesn't just happen through skill-building; it happens through sponsorship, when a leader mentions your name in a room you aren't in yet. * Controlled Imbalance: Rejecting the myth of "work-life balance," Hui shared her practice of "controlled imbalance" - hard-blocking 5 PM – 9 PM for family and being unapologetic about it. Special thanks to Westin Sherlock Jing Xue Robert King and all those who contributed to a lively discussion! Spending this time with Dr. Hui anchored our Women’s History Month celebrations in a way that felt both timely and profoundly authentic, for all of use here at Signifyd. Thank you so much, Hui!

98

3 Comments

I recently contributed to CLI-Anything, a framework by the HKU HKUDS team that makes desktop software controllable by AI agents. The problem: AI coding agents like OpenClaw, Claude Code are great at writing code, but they can't operate GUI applications like Gimp, Blender, or MuseScore. These tools may have CLIs, but they were designed for scripts — cryptic flags, undocumented JSON schemas, noisy output. Not something an AI agent can reliably work with. The solution: CLI-Anything wraps each software's native CLI into a standardised, agent-friendly interface — consistent command structure, clean JSON output, discoverable --help on every command, and stateful REPL sessions with undo/redo. My contribution: I built the MuseScore 4 wrapper — the first music notation tool in the collection 🎵  It enables AI agents to:  - Transpose scores between keys  - Export to PDF, MP3, MIDI, MusicXML, PNG  - Manage individual instrument parts  - Analyse score metadata and statistics  - Track changes with undo/redo sessions A real-world use case: a piano teacher can run a single command to transpose a piece from G major to C major for a beginner student, then export a PDF for printing and an MP3 backing track for practice — all without opening the GUI. The project now covers 16+ applications across image editing, video production, 3D modelling, audio, and now music notation.  🔗 GitHub: https://lnkd.in/gUh5JbFS  🌐 CLI Hub: https://lnkd.in/gK_PDxt3  #OpenSource #AI #MusicTech #MuseScore #AIAgents

19

BECOME AN INDIVIDUAL CONTRIBUTOR Over the past few months, I've seen several friends and ex-colleagues proactively move from leadership (manager or managers-of-managers) roles at non-AI companies to IC (individual contributo) roles at AI-native companies. While this might seem counter-intuitive and maybe even a career negative, I believe it's the right career move for leaders who want to move into AI. Every AI-native company is looking for leaders with AI native experience. If someone stays a mid-level manager in a non-AI role, they will not build the hands-on AI experience they need to become an effective AI leader. Even a couple years of IC work in an AI native environment will give them confidence and the optionality to either continue down the IC route or go back down the managerial path, now armed with real AI building experience and so a much better manager. Many AI companies are growing so fast that you might even get a "battlefield" promotion to people manager sooner than later, esp if you have prior management experience. Leaders / people managers in non-AI roles: If you're feeling stuck, please consider interviewing for and moving into an IC role at an AI company. It'll be lots of fun, plus you'll build durable skills you will use for the next several decades. PS: Do it fast. Don't wait too long. The half life of non-AI roles is shrinking fast.

2,236

156 Comments

How did LangTalks end up meeting with Andrew Ng, Anshul Ramachandran (Windsurf), and Roey Zalta And what book does #LarryPage wish he had when he started Google? It all relates to the question that keeps me up at night: how should our evolving understanding of 'search & retrieval', 'context engineering', and 'evals' shape the way we design systems? And how should we account for these shifting dynamics in the planning stage with a forward looking approach? (Though Page’s recommendation is real, so I hope you read until the end.) First, the new course Agentic AI with Andrew Ng is one of the best resources I’ve found (especially Module 4 on evals) [1]. I also recommend listening to LangTalk’s episode on search & retrieval with Guy Itach [2]. I especially liked the ending, where they show how technical aspects are combined with product sense and domain expertise, and why choosing the right evals isn’t straightforward at all. Don’t miss Roy Zalta’s post [3] on how he achieved 98% accuracy in his RAG system using a combination of eleven different strategies. And last but not least: Windsurf’s AI Coding Agents [4]. Specifically, the section on “Search & Discovery for AI Agents,” where Anshul presents a multi-step retrieval paradigm—that blew my mind (it’s free). Another thought I’ve had on 'evals and error analysis' is that it's somewhat derivative of the OKRs methodology—an incredibly powerful concept for driving clarity, prioritization, organizational alignment, and decision-making. I honestly can’t recommend Measure What Matters [5] enough—but don’t take my word for it. See what Larry Page, Alphabet CEO and Google cofounder, had to say: "I wish I had had this book nineteen years ago, when we founded Google. Or even before that, when I was only managing myself! As much as I hate process, good ideas with great execution are how you make magic. And that’s where OKRs come in." Circling back to my question: As we shift toward agentic systems, how do you see the evolution of system design? What planning strategies are you starting to prioritize from the very beginning?

6

1 Comment

The energy at our recent Google DeepMind AI Hackathon in Korea was nothing short of electric, proving once again that Korea is absolutely the place to be for developer innovation 🇰🇷 🏆 Congratulations to Winners! 1st place: Minsu Chang for refining spatial mapping for mobile robots 2nd place: YongGyu Kim for buiding an interative gamification of AI security protocols 3rd place: Jaeho Shin for building a rapid AI assisted Webtoon creation The level of "Vibe Coding" and technical depth we saw proves that the Korean developer community is leading the charge in the era of Agentic AI. Thank you to everyone who joined us and we look forward to engaging with you again soon! NaHyun Kwon Amit Vadi Joana Carrasqueira Ju-yeong Ji Jeongkyu Shin Jaeyeon Baek JeongMin Kwon Junbum Lee PARK JAICHANG 한성민 Minho Ryu 변성윤 Sungyong An Omar Sanseviero Manikantan Krishnamurthy Sami Kizilbash SUKONE HONG Sierra Okazaki Adam Blumenfeld Mike Lee Ray Del Vecchio Haimin Lee

35

6 Comments

During my 15 days in China, one thing stood out everywhere AI adoption is moving at lightning speed. From Chengdu to Shanghai, Beijing, Shenzhen, and Chongqing, startups and big companies are integrating AI into everything, especially high-tech hardware and next-gen solutions. At ByteDance, I got a glimpse of how AI is powering content, automation, and foundation models, creating products that scale globally from day one. The level of speed and ambition here is something I rarely see in India. For Indian founders, the takeaway is clear. Think bigger, move faster, and don’t be afraid to build high-tech solutions that aim for global impact. China’s ecosystem proves that AI is not just a tool, it’s a foundation for building the next generation of scalable businesses. For QuickReel it’s a reminder to always bet on big, impactful technology.

18

SK Hynix Accelerates HBM Supply and Turns AI Memory Into an Ops Race The AI buildout is now gated by HBM stacks and packaging throughput. When fabs move three months early, your forecast model needs updating. ‎ 📊 2026 Execution Performance Snapshot • M15X starts production in February 2026, pulled forward by about three months. • Initial ramp targets 10,000 wafers per month, with expansion through 2026. • The first Yongin fab moves to February 2027, also pulled forward by three months. • The Yongin cluster scales to four mega fabs under a 600 trillion won plan. • HBM4 ramps for Nvidia Rubin, with HBM4E stacking aimed at 2027 volumes. ‎ 🧭 Commercial Strategy Levers • Treat HBM as the control plane for AI infrastructure delivery, not a commodity. • Pull forward wafer starts to lock allocations while rivals chase qualification. • Invest in packaging early, since bonding and stacking now set the real limit. • Use scarcity to defend pricing power as data centers and devices compete. • Convert capex into multi-year share through repeatable ramp and yield learning. ‎ 👔 For Operations, Revenue, and Growth Strategy Leaders • Model AI rollout risk across wafers, packaging, yields, and customer commit dates. • Build governance around allocation, not only procurement savings. • Forecast hardware revenue with delivery cadence. • Treat advanced packaging as a capacity plan, with its own KPIs and owners. • Pressure-test “AI ready” claims against physical constraints and lead times. ‎ 💬 Curious how to model HBM supply risk in AI rollout plans? Let’s connect. ‎ ‎▪ Disclaimer: Personal views only. Not financial, or investment advice. #SKHynix #HBM #AIMemory #Semiconductors #AIInfrastructure

8