Joseph Zhong

Lemurian Labs just closed a $28M Series A that feels less like a funding round and more like a pressure wave moving through the AI infra world. pebblebed Ventures and Hexagon co-led the round, and that pairing does not usually appear unless something fundamentally shifts under the surface. Oval Park Capital returned after leading the 2022 seed round, linking arms with Untapped Ventures, Silicon Catalyst Ventures, Origin Ventures, Blackhorn Ventures, Uncorrelated Ventures, 1Flourish Ventures, Planetary Ventures, Animal Capital, and Stepchange Ventures. When that many firms with that much pattern recognition converge, they are not buying hype. They are buying inevitability. The inevitability comes from a simple but brutal truth. AI infra is bloated, fragmented, and expensive, and every #developer who has ever rewritten the same kernel for the 3rd time in a month knows the pain. Jay Dawani founded Lemurian Labs to kill that pain, not with a mascot or a mission statement, but with math that refuses to blink. Jay teamed with Dr. Vassil Dimitrov, whose work on compilers and #GPU systems has quietly powered more of #moderncomputing than most people realize. Add CTO Theodore Omtzigt, VP Engineering Christopher Vick, VP Product Abhay Chitral, Head of AI Infrastructure Adam Robertson, Chief of Staff Jenny C., and Head of People Ops Bryan Carpender, and suddenly the company looks like someone reverse engineered a dream team from Sun, Intel, NVIDIA, Qualcomm, Altera, IBM, Uber, Google, and then stitched it together under one roof. The centerpiece is Tachyon, the company's universal stack that lets developers write #AIcode once and run it across NVIDIA, AMD, Intel Corporation, #CPUs, #NPUs, and edge devices without losing performance. No vendor lock-in, no late night rewrites, no praying the hardware gods behave. The compiler decomposes models into task graphs, maps them to any #ISA, generates optimized object code, and hands the runtime a system that keeps tuning itself as workloads shift. The twist comes from PAL, the Parallel Adaptive Logarithms number system that turns multiply-divide operations into add-subtract motions, the computational equivalent of replacing a gas guzzler with a precision-tuned electric motor. AI workloads are on track to swallow 20% of global electricity by 2030-2035, and inference costs are ballooning so fast that #hyperscalers are starting to look like power companies with side hustles. A platform that can cut those costs by 60-80% while improving hardware utilization is not a tool. It is leverage. With early partner access this year and a beta landing in summer 2026, Lemurian Labs is positioning itself exactly where the market's pressure is highest. This Series A is fuel, but the direction was already set. #Startups #StartupFunding #VentureCapital #SeriesA #AI #Infrastructure #AIInfra #DevTools #Coding #Compute #Technology #Innovation #TechEcosystem #StartupEcosystem

26

3 Comments

A great article by Bryant Francis on how game developers are getting more worried about generative AI lowering the quality of games: https://lnkd.in/efNg6tX5 I'm quoted in the piece, and here are the complete thoughts I sent over when asked to think about the question. TLDR; yeah. But it's messy. When people talk about Generative AI and quality, it's because these models output content that resembles the median of what they see in the training data. That means that LLMs are **aspirationally** mid, and often worse. It's my experience that game developers are a passionate crew, driven to make something creative and new and amazing; this is not what Generative AI does. Couple the technology with the economic pressures of the games business, and there will definitely be folks using Generative AI to crank out mediocre or even slop games without the care and polish that a human game developer could bring. It's no surprise that people who care deeply about the craft of game design are not excited about this. The reality is a bit nuanced, and a couple of things will happen; new kinds of games will be possible, and we're in a wide open moment to create those now, and we'll develop better systems to get through the slop. Even today, the technology can be tremendously useful in the right hands as an accelerant to experimentation or in a well-understood process with great quality controls, and beyond that, the real potential is in enabling games and experiences that couldn't possibly exist without it. As one example, at Hidden Door we create a storytelling experience for fans as close to having your favorite author as an improvisational GM as possible -- an old game style, now possible in a new form. I'm also not convinced that the slop will matter; it shifts the burden onto the distribution platforms to help players find the games they want to play, and onto designers to create features to help bring players together. These are the same UX problems faced by all UGC platforms, which opens up an opportunity space for new approaches to distribution, and can make the economics more interesting, too.

31

6 Comments

Working with LLMs to create prototype webapps from my phone can be very satisfying. There are limitations, though. I only see snippets of code as they fly by. I ported the Microprints library that Stephan Eggermont made to live inside our web application. This way I can see the forest and some of the trees :-). I now can work on intent (prompts), user experience (mobile first) and keep an eye on the code - what evolves now? how does it look in detail? Do I need to stop and rethink my intent or design, refactor code or add tests? Link to blog post in the first comment. #MoldableDevelopment. FYI #Elixir and #Phoenix #LiveView make this easy (with a coding agent). Phoenix LiveDashboard has a plug in system, and claude code knew how to use it, the 'Dev Tools' here are just one of the tabs in the Live Dashboard (which is a nice starting point for #Observability).

6

1 Comment

Anthropic just blocked third-party use of Claude Code subscriptions. Sound familiar? In 2023, Apollo did the exact same thing. Shifted their license from open source to a restrictive source-available model. Pulled the rug on their community. That decision created WunderGraph's opportunity. After years of working with enterprises like eBay, Paramount, FanDuel, and others, here's what I've learned: Open source will always win in infrastructure. Why? Enterprises are complex. They want control. They want customization. They want to audit the code running their critical systems. With WunderGraph Cosmo being fully open source, our customers can: - Audit every line of code - Extend the platform for their specific needs - Contribute back to the codebase - Run it anywhere they want No license changes. No vendor lock-in surprises. No "we changed the terms" emails. Every time a company restricts their license or usage, they hand their competitors a gift. Apollo handed us one in 2023. Anthropic just handed someone else one today. Our goal is simple: if you're running GraphQL Federation, you're runnin WunderGraph. Open source makes that possible.

33

4 Comments

I used Claude Artifacts to create a website to observe AI collaboration - a system where two LLMs work together to solve complex problems, with one acting as a quality checker. The Challenge -> Build a web application that would: ->Take user problem statements as input -> Have one LLM (SOLVER) generate solutions -> Have a second LLM (QUALITY) act as a demanding quality assurance expert, but providing specific guidance on how to make it better -> Create an iterative feedback loop (maintain context) until the solution meets extremely high standards -> Output the entire collaboration process + final approved solution The first time I tried this Claude was giving me gibberish, but it sort of worked today! My main objective was to observe the agents collaborate and go back-n-forth in real-time Key Insights -> SOLVER made up facts, and produced really bad results in the first few iterations. I tried a couple of different problem statements - one around Chess, and another about Steph Curry’s impact on Basketball. Here are some issues identified by QUALITY, and verbatim feedback for SOLVER: -- FEEDBACK: This solution contains multiple critical errors that make it fundamentally flawed and unusable: -- ..This is nonsensical .. -- You claim "Re8#" is checkmate but then immediately show Black playing "Rxe8" - if Black can capture the rook, it's not checkmate. This is a basic chess rule violation. -- “Provide concrete statistics, expert quotes, specific examples, and address the analytical gaps before resubmission” -- **INCOMPLETE SECTIONS:** ---- The "Rule and Officiating Adaptations" section cuts off mid-sentence with no content provided -> The quality of output significantly improved through this adversarial collaboration -> Often, the collaboration would reach the max iteration limit of eight! -> CHECKR's role as a demanding reviewer prevented mediocre (if not inaccurate) solutions from passing through -> Maintaining conversation context was crucial for meaningful iterations -> I can’t say if I still trust the final solution! If SOLVER can make errors, so can QUALITY (?) Exciting nonetheless!

12

Anysphere (Cursor), that I profiled recently appears to be launching Composer 2. The shocking part? They claim it’s a frontier-level coding model that matches or beats the performance of flagship models like GPT-5.4 and Opus 4.6 at roughly 1/10th the cost. Composer 2 scores 60.0 on CursorBench and 59.8 on Terminal-Bench 2.0, outperforming Opus 4.6 and Sonnet 4.5. Cursor has priced the model at $0.50/M input and $2.50/M output tokens - drastically lower than any comparable frontier model. Launched a new RL process called self-summarization that allows the model to master complex tasks requiring hundreds of individual actions. Bloomberg and The New Stack also have pieces out on this today.

7

5 Comments