Lev N.

Cursor's new Composer model is (not so quietly) Chinese under the hood. Cursor shipped the new version of Composer as its "first in-house frontier coding model," optimized through RL for real coding tasks. Within days, users noticed Composer's hidden thinking was full of Chinese characters. Industry press connected the dots: both Cursor's Composer and Windsurf's SWE-1.5 appear based on Chinese large models (likely Qwen bases), fine-tuned and wrapped in US products. Community reaction is split. Some call it a major downgrade: slower and struggling with complexity. Others report speed gains and slightly better performance. So why is the reception so mixed? Gavin Leech's AI Tigers post (link in comments) shows that on fresh benchmarks, Chinese models lose around 21% performance compared to roughly 10% for Western models. They look closer on leaderboards than they perform on novel problems. The cost story is similar: 3-6x cheaper per token, but often requiring 2-4x more tokens, which shrinks the cost advantage fast. So why would Cursor make this trade? Because they're not trying to replace GPT or Claude. They're owning a narrow, high-volume behavior and squeezing it with fine-tuning. Cursor has massive amounts of product-specific data (traces, diffs, accept/reject signals) and can invest in the RL loop to extract value from a cheaper base. This is where Chinese and open-source models actually fit in the stack. They're a strong fit for narrow problems with abundant training signal: tab complete, inline edits, code transforms. They're a weaker fit as primary reasoning for high-stakes, long-horizon workflows. Frontier US models still dominate the second category. Chinese open-source is taking over the first. Composer's rollout is a case study in how powerful Chinese models have become, and a live demo of where their limits still show.

39

12 Comments

I’ve just published the next installment in my "Theory of Durable Execution" series. Across frameworks we hear terms like workflows, activities, virtual objects, handlers, and functions. Underneath the terminology, they’re all describing a small number of execution patterns. In this new post, I break durable functions down into three forms: 🔹 Stateless functions for one-shot deterministic logic 🔹 Sessions for long-running, interactive orchestration 🔹 Actors for persistent, message-driven state To get there, I walk through: 🔹 The behavior–state continuum 🔹 How identity, lifetime, communication, and concurrency shape execution semantics 🔹 How frameworks like Temporal, Restate, DBOS, and Resonate map onto these forms This builds on the previous two posts in the series: 🔹 Demystifying Determinism in Durable Execution 🔹 The Durable Function Tree (part 1 and 2) This was the last post in this series, distilling the various frameworks down to a simpler logical model. Full post: The Three Durable Function Forms https://lnkd.in/eJHpY9qC

77

7 Comments

🤓 Woot, woot, very cool to see: Folks giving #Hardwood a try for parsing their Parquet files. Performance hasn't been a big focus yet, but it seems good enough when processing entire data sets (thus not requiring predicate push down). 👉 https://lnkd.in/eaty-Q6g

23

Interested in learning more about Foundry Local https://foundrylocal.ai and how to build RAG-style systems that run locally? I’ve been exploring patterns for local-first retrieval, generation, and agent workflows, focusing on privacy, performance, and developer control. If you’re experimenting with RAG beyond cloud-only setups, these samples might be useful: • Local RAG – a simple, practical starting point for retrieval‑augmented generation running locally https://lnkd.in/e8DaV8ae • Local CAG – moving beyond classic RAG into contextual and agent‑driven generation patterns https://lnkd.in/eb5u_QXc • Local Hybrid Retrieval (ONNX) – combining lexical + vector search with optimized local inference https://lnkd.in/eh4mMwSi These projects are all about understanding the building blocks: local models, embeddings, hybrid retrieval, and how they fit into modern agent architectures with Foundry Local. If you’re curious about edge AI, local-first agents, or privacy-aware RAG, feel free to explore, fork, or share feedback. #AI #RAG #FoundryLocal #AgenticAI #LocalAI #EdgeAI #Developers

18

📝 Wrote a thing: "'Streaming vs. Batch' Is a Wrong Dichotomy, and I Think It's Confusing" Oftentimes, Stream vs. Batch is discussed as if it’s one or the other, but to me this does not make that much sense really. They actually go together, and the more important distinction to make is Push vs. Pull. 👉 Read the full post here: https://lnkd.in/eNsq5DYq

136

10 Comments

Kagi has a translator that will turn English into McKinsey-speak. Also Gen Z. The avian asset’s presence facilitated a positive shift in emotional sentiment, driven by its high-gravity aesthetic and disciplined posture. I initiated a stakeholder inquiry: "Despite your suboptimal grooming and lack of plumage, you demonstrate significant risk appetite. As a legacy entity from the nocturnal ecosystem, what is your core brand identity on the Plutonian shore?" The Raven responded with a definitive, non-negotiable strategic pivot: "Nevermore." https://lnkd.in/ezKWuRxQ

1

Boris Cherny’s View on Vibe Coding Limitations (from the Podcast) • Best for throwaway/prototype code: “I do this all the time [vibe coding], but it’s definitely not the thing you want to do all the time.” It’s ideal for non-critical path code, quick experiments, or prototypes that might get discarded. • Lacks maintainability and thoughtfulness: “You want maintainable code sometimes. You want to be very thoughtful about every line sometimes.” AI can produce working code fast, but it often results in messy, hard-to-maintain output (e.g., inconsistent style, over-engineering, or subtle issues). • Requires human oversight for quality: On the Claude Code team, they hold AI-generated code to the exact same bar as human-written code. If it “sucks,” they don’t merge it—they ask the model to improve it. • Preferred approach: Pair programming with the model: For important code, align on a plan first (e.g., using Claude Code’s plan mode), iterate incrementally, review/clean up, or even hand-write critical parts. Boris still manually codes core sections where he has strong opinions (e.g., parameter names). • Models aren’t perfect yet: “The models are still overall not great at coding… this is the worst it’s ever going to be.” Massive improvements are coming, but current limitations mean vibe coding alone isn’t reliable for production. Broader Limitations of Vibe Coding (from Industry Discussion & Karpathy’s Own Reflections) Karpathy popularized the term but has since highlighted drawbacks, even abandoning pure vibe coding for complex projects due to persistent bugs and “slop”: • Produces “slop” or low-quality code — Hallucinations, duplicated logic, bloated/over-engineered solutions, inconsistent architecture → Leads to technical debt and hard-to-debug systems. • Security vulnerabilities — AI can introduce exploits (e.g., insecure dependencies, leaked keys) that humans might miss if not reviewing deeply. • Hard to debug and maintain long-term — Lack of deep understanding means subtle/non-obvious problems pile up; fixing requires trial-and-error. • Not suitable for production/critical software — Fine for weekend projects or demos, but risky for anything with real stakes (e.g., scalability, reliability). • Paradox: Best for experienced coders — Experts can spot/fix issues; novices risk building fragile apps without realizing. In summary, vibe coding is a powerful tool for speed and creativity (especially prototypes), but its limitations make it unsuitable as a full replacement for thoughtful, reviewed coding—especially in professional settings. As Boris puts it, treat AI like a pairing partner: collaborate actively, enforce high standards, and intervene where needed. Models will get better rapidly, narrowing these gaps over the next months/years. https://lnkd.in/e446EFR3

4

Language mixing isn’t just a quirky bug — it’s a real challenge for multilingual AI. This article explains why LLMs slip between languages and how to reduce it in production. https://lnkd.in/g-jpTUdF #llm #nlp #deeplearning #languagemixing

20

This is absolutely true. I would also add that just because AI decreases cycle time does not mean you should operate without a plan. To use an analogy, pretend you are going on a trip from New York to Los Angeles and fuel/electricity are free (or only cost pennies). Now imagine that you don't have a map but are just driving based on instinct. Because you are moving and the cost is low or zero, you feel you are getting somewhere even when you are not. You need a map (app or paper), or at least need to stop for directions occasionally. Movement, no matter the speed, is not progress unless it's toward where you want to go.

3

2 Comments

OK. Before I get into this, a quick disclaimer: This is my personal opinion as an individual. It does not reflect the views or positions of Adobe in any way. Nothing here constitutes financial or legal advice. These are simply my own thoughts. There is a lot of speculation right now about Adobe and generative AI. The market narrative says AI will disrupt creative software and undermine the subscription model companies like Adobe built over the last decade. Yet Adobe continues to report strong growth. The stock still dropped because of broader fears about AI competition and leadership changes. I think the market has this one wrong. Many analysts still wrongly frame Adobe as a consumer creative software company. In reality, Adobe runs enterprise-scale ecosystems across Creative Cloud, Document Cloud, and Experience Cloud that power creative production, marketing operations, analytics, and document workflows for major organizations worldwide. Enterprise adoption tells a very different story from the market narrative. Nearly every Fortune 100 & 500 company uses AI features in Adobe apps, and most large enterprise customers have already adopted AI-driven tools like Firefly services, GenStudio, and AI assistants. This is not novelty AI. Consumer AI tools often work like slot machines. You type a prompt and hope something good enough appears. Enterprise creative production requires governance, repeatability, brand control, and legal safety. That is where Adobe operates. Generative AI can produce generic content quickly. It does not eliminate the need for creatives. If anything, it increases the demand for people who understand storytelling, design systems, and brand vision. Creative work has never been about pressing a button. It has always been about vision. I believe Adobe’s strategy reflects that. AI is being embedded directly into creative workflows so professionals can move faster while keeping control of the output. From where I sit, this is a case of analysts underestimating the value of creative professionals and the systems they rely on, and ignoring business offerings. Folks who are not creative can't make calls about how creatives work and what their needs are, and I don't think many startups understand enterprise needs at scale or what Adobe's ecosystem brings. I work with these folks, I have a different perspective. I believe the market is undervaluing the platform. I am still investing. Maybe I am wrong, maybe I am right. But if you want to make an omelette, you have to crack a few eggs. But there's one thing I do believe, in a world where AI agents are deleting years' worth of databases, self-driving cars can be captured in what amounts to painted salt circles, computer vision is targeting pictures of planes on the ground, and companies having major security issues in the software AI writes... I don't trust a computer with the most important parts of my business, only the busywork. And busywork has never been the vision.

35

6 Comments

This week, Anthropic's Claude 4 seems to be again on a roller coaster: after pushing Amazon's new Kiro IDE (which was briefly available for download before an avalanche of demand slammed it shut), his coding performance plummeted noticeably. Tool usage has dropped. The reason? Degraded. And as usual at Anthropic, no one knows why. We just get silence and a fresh dose of usage limits. Meanwhile, Microsoft's Copilot is getting a makeover literally. The new "appearance" feature could soon allow it to simulate traits like age or personality. Whether it's charming or dystopian depends on your tolerance for digital companions with synthetic charm. But one thing is clear: AI is no longer just smart, it's present. Speaking of IAG, Meta appointed Shengjia Zhao to lead its new Superintelligence Lab. If that name doesn't ring a bell, it soon will: Meta is delving into fundamental models with AI ambitions, indicating that the "big three" (Meta, OpenAI, Google) are gearing up for the final showdown. But it's not all polished demos and ambitious projects. A troubling moment at Sketch.dev exposed the dark side of AI programming assistants: a subtle AI-written database optimization caused a crash in production under load. Reminder: just because your AI can write code doesn't mean it should. Test, test, and test again. Mistral's Codestral also launched, quietly stealing the show. With 22 billion parameters and support for over 80 languages, it's open-source, meaning it has no barriers. Elsewhere, Google's NotebookLM received a major update powered by Gemini. It now summarizes YouTube videos and drafts of content like your personal caffeine research companion. Think of it as a fusion of CliffsNotes and ChatGPT: optimized, hyper-efficient, and incredibly good at synthesizing ideas. In the worst-case scenario for cybersecurity, researchers have revealed AI-powered malware that thinks for itself. It interprets its environment, adapts, and executes evasive strategies with chilling precision. It's not HAL yet, but it's time we reconsider the true meaning of "autonomous threat." Finally, some news that caught my attention: A new concept unlike the transformers and tokens we know today, the Hierarchical Reasoning Model (HRM) concept promises to make the large models we know more efficient. In its first milestone, a small model of just 27M was able to beat o3-mini on pre-trained tasks. for more information follow me here: https://lnkd.in/gNVYeZST

4

Advantages of Working Collaboratively: In our increasingly complex world, the secret power often lies not in singular genius—but in collective intelligence. As Colin M. Fisher writes in The Collective Edge: “The secret to getting the most from your groups is knowing how to work with the invisible forces of group dynamics instead of being mindlessly pushed around by them.” Here are a few key advantages of working collaboratively, inspired by Fisher’s insights: 1. Synergy & Creativity When structured well, groups can be more than the sum of their parts. Diverse perspectives spark innovation, as Fisher illustrates with design firms and high-performing teams. 2. Healthy Competition Competition within a collaborative group doesn’t have to be destructive. It can motivate without leading to conflict, if the group norms support it. 3. Constructive Conformity Conformity isn’t always a bad thing — under the right conditions, it can promote alignment, shared goals, and group cohesion. 4. Effective Leadership through Structure Great leadership doesn’t always mean charismatic speeches; sometimes it's about setting the right structure, coaching, and creating psychological safety. 5. Better Decision Making By paying attention to group-level dynamics, teams can make more robust decisions — avoiding the pitfalls of blind conformity or domination by a single voice. — When we lean into the hidden dynamics of groups, we unlock potential that individual effort alone can't reach. Let's celebrate and build collective edge in our teams. #Collaboration #Teamwork #GroupDynamics #Leadership #Innovation #CollectiveIntelligence #WorkTogether #CollectiveEdge

2

One of my favourite WebGPU demos is the "deferred rendering" demo (or as I call it, "disco dragon"). It's pretty, and it demonstrates a cool rendering technique for highly complex scenes with lots of dynamic lights: https://lnkd.in/eVE226HN So I rewrote it in Toucan, my graphics-oriented programming language: https://lnkd.in/e2Kwe334 (runs in Chrome and Firefox Nightly). It's 424 lines of Toucan, versus 801 total lines of WGSL and TypeScript. Most of the savings come on the host side, where, having defined our vertex buffers, uniform buffers, bind groups and pipeline variables once for the device side, we don't need to re-declare them for the host side in a new syntax. Plus, a common type system means there's no need to compute byte offsets or vertex strides, match up group numbers or bindings numbers, re-declare attachments, or accommodate std140 or std430 memory layouts. All of that is done automatically by the compiler. Type safety means not only being able to assign to members of a vertex or uniform buffer on the host using the same syntax as the device, but also applies to the usage itself: since the "vertex" and "uniform" qualifiers are understood by the host side, they become part of the type. If a required qualifier is omitted on allocation, attempting to bind the buffer as such will result in a compile-time error, not a runtime one. Here's a diff of the two versions (with comments and whitespace added to help the diffs to match): https://lnkd.in/ew64MGGS

8

7 Comments

Open Responses is an open-source spec based off OpenAI's Responses API, with backing so far from OpenRouter, Vercel, HuggingFace, LM Studio, Ollama, and vLLM. https://openresponses.org Will more providers get on board? It'd be so nice to have a true standard for LLM APIs.

69

5 Comments

The most succinct way to put how AI is extracting value from creatives and humans in general without giving anything back .

1

As COO at www.jaiinfoway.com this is an important signal for where agentic AI is heading. Non technical implementations are catching up fast but the real story is iteration speed and learning loops. Building Cowork entirely with Claude Code shows how agents are becoming first class workers not just assistants. At www.jaiinfoway.com we see this shift clearly as enterprises move from copilots to autonomous task execution with guardrails. UX will evolve tooling will mature but the direction is set. Curious how others here see non technical users adopting agent driven workflows this year. Are we ready for agents as coworkers. #Jaiinfoway #AgenticAI #Claude #AIAgents #EnterpriseAI #FutureOfWork #AIProductivity #AITransformation #AILeadership

2

In this blog Mohammed Rafiq and I dive into the technical details of how Ragie's audio video pipeline and retrievals work. We share some of the approaches we considered and the reasoning behind the design we landed on. If you get a minute, give it a read and let us know what you think.

6