Sean Helvey

A developer in San Francisco selects MiniMax as her inference backend. The request crosses the Pacific, gets processed in a Chinese data center, returns in a second. She pays $0.30 per million tokens. Claude Opus would cost $5.00 for the same volume. No customs declaration. No tariffs. No entry in any trade database. Chinese tech media calls this Token 出海: Token Export. This month’s OpenRouter data showed it crossing a threshold: Chinese models surpassed American ones in total token consumption for the first time. This did not happen overnight. I spent February tracing the structural forces behind it: efficiency architectures, energy cost gaps, and a competitive intensity that pushes pricing below rational margins. Nvidia’s record quarter and falling stock price are part of the same story. China’s Compute Bet explains the system. https://lnkd.in/ghEBwk-s The Export That Tariffs Can’t Touch explains what happens when that system meets the global market. https://lnkd.in/gh_ajmmZ

3

🧑‍💻 Evaluating the new Gemini 2.5 Pro update on R coding ✨ Last month, Simon P. Couch assessed Gemini 2.5 Pro's capabilities in R coding tasks. Now, with Google's recent update to Gemini 2.5 Pro, Simon revisits the evaluation to see if there's a notable improvement. In this follow-up, he compares the updated Gemini 2.5 Pro against its previous version, as well as leading models like Claude 3.7 Sonnet (with thinking enabled) and GPT o4-mini. Using the {vitals} package and the 'An R Eval' dataset, Simon provides a detailed analysis of each model's performance. 📌 Key takeaways: - The updated Gemini 2.5 Pro performs similarly to its previous version in R coding tasks. - Claude 3.7 Sonnet (with thinking enabled) and GPT o4-mini lead in this evaluation. For a comprehensive breakdown of the evaluation and insights into each model's performance, read Simon's full post, below. #RStats #AI

6

If corporate data can’t move to the cloud, LEANN is a practical way to ship semantic search + RAG fully locally.​ MIT license, open source. Enjoy 97% storage savings while running a fast, accurate, and 100% private RAG application on your personal device. https://lnkd.in/daMD4JAZ It’s designed to index and search very large corpora while using far less storage by recomputing embeddings on demand instead of storing everything.​ Where LEANN fits - Use it when security, compliance, or contracts require data to stay on device / on‑prem.​ - It’s built for “personal AI” / private RAG across many sources while keeping data local.​ Practical use cases - Internal policy + SOP Q&A (HR, Legal, Security) without uploading documents.​ - Engineering knowledge search across codebases and docs (semantic search vs keyword-only workflows).​ - Private “ask my files” assistant for teams handling regulated data (finance, healthcare, defense suppliers).​ - Local investigation / eDiscovery-style retrieval on approved archives (docs, chats, exports) with minimal storage overhead.​ Why teams care - Privacy-first: data stays local; no forced cloud dependency.​ - Lightweight indexing: targets major storage savings versus traditional vector DB approaches.​ - Flexible generation: supports local inference engines (Ollama recommended) and OpenAI-compatible endpoints when your environment allows it.​ Biggest blocker for your org for internal RAG? Data residency, cost, or latency? #AI #RAG #EnterpriseAI #OnPrem #DataPrivacy #DataGovernance #InfoSec

1

https://lnkd.in/gUK5W5YT  Truly open source AI isn't quite where it needs to be. Couple that with the expense of GPUs/hardware and a true "democratized" and private LLM will be difficult. As in all new tech, private AI will start with the enterprise and work its way down.

4

Reflections from 1 week in SF: - Dev tool space has never been hotter (emphasized by Figma's recent IPO!) - Agentic UIs and workflows are becoming the bare minimum - Fundraising is rarely a concern to founders in the Bay area. Founders are raising pre-seed with a Notion doc + hype (despite macro sentiment!) - Talent density is unparalleled (so much so that I met a cracked engineer living in his car just to be close to the Bay area) - Most founders I met were both extremely technical and could sell ice to eskimo - Vertical AI platforms could be huge disrupters to mom+pop shops….Think AI-native ERPs for niche markets: AI for insurance underwriting, AI-first factory optimization, AI enabled supply chain ordering for dentists etc. Verticalised solutions with a full-stack GTM, co-built by domain experts, with a network as a moat = 🔥 - Founders are truly pushing frontiers of innovation more than I've ever seen before. Everything from agentic deployment in <5 mins, agents working with agents, agentic browsers [see the trend?], fitness non (!) wearables, poop scanners, chat to 3D print & more - Tremendous long term efforts in bringing together the ecosystem for upskilling, knowledge sharing & connecting. Literally attended startup events held at a library, a community centre, a comedy club, an accelerator & the AWS's builders loft (shoutout Amazon Web Services (AWS) for all the free food 😂 ) - AI on billboards everywhere (see pic!)

46

5 Comments

Is RAG Dying? Not Quite, But It’s Not Always Needed. When building an agentic immigration solution, I started with what many consider best practice: RAG. We used a vector database hosted on Vertex AI, passed embeddings to a smaller Gemini model to summarize context, then fed that to Gemini 2.5 for action. The outcome? Around 80% accuracy. Then came the good old approach. F&&k the what theory says and try it differently, so we instead saved full documentation in a Postgres database. An orchestrator agent translated user intent into SQL, pulled relevant documents from S3, and passed them as-is to Gemini 2.5. The result? Nearly 100% accuracy. Far fewer edge-case issues. And get this—1/4 the cost compared to the RAG setup. This wasn’t theoretical. It’s weeks of live, production-grade testing. RAG still has its place, especially when your data is vast and unstructured. But don’t assume it’s the answer for every use case. Sometimes, a clean, structured pipeline and direct document access are all your model needs. If you're evaluating RAG or considering an agent-based flow, let’s chat. I’d love to brainstorm and share more insights.

31

1 Comment

If you've been using Claude Code, you already know this: agents are becoming the primary consumers of software. Need a database? ask the agent. Need auth? let it handle it. Need to enrich a lead list? you got the idea. Maybe you have a preference, but if you don't specify it, the agents choose for you. Every software company becomes an API. And with that comes a problem few people are talking about: new criteria to compete for agentic selection at runtime. Most companies have no idea how they rank on those criteria, let alone how to improve them. No traces, no signal. No visibility into what an agent tries first or where it gives up. The vendor's dashboard shows zero errors. The agent already moved on.

14

1 Comment

Had a good/long talk with the team about vibe coding. Keep in mind that LLMs don't hallucinate 5% of the time. They hallucinate 100% of the time, we just agree with them 95% of the time. Without a human reviewer you are just asking for trouble. Think of it as hiring absolutely whip-smart junior devs, who are insanely fast at cutting and pasting from Stack Overflow. That's good, except they don't understand your business, business processes, security needs, etc... etc. So if you don't review the code, no one will. h/t to Casey Ellis for the graphic idea.

42

9 Comments

One of the biggest problems we see with agent systems today: They don’t remember anything useful. Agents restart every session with no durable context 🧠 So we built the Memori Labs OpenClaw plugin to change that - giving OpenClaw gateways real long-term memory across agents. Excited to see what builders create with it 🚀 https://lnkd.in/eEvnHs9x

28

2 Comments

🔥 Curiosity about useState drove me into the world of Fiber and reconciliation I started learning about the useState hook and how the React team actually built this thing. It looks like everything is tightly connected with Fiber. The hook logic lives inside Fiber, and the whole idea of custom hooks is just to reuse logic anywhere you want. The only rule is the function name must start with use*, but inside we just write normal JavaScript logic. When I checked how useState is implemented, I noticed they use Fiber heavily. Then I fell into the rabbit hole of Fiber and reconciliation. Very interesting stuff. Before React 16, they used something called stack reconciliation. It was fully synchronous, no priorities, no way to pause, and everything executed through the JavaScript call stack (LIFO), which caused a lot of problems for big apps. Then Fiber came and completely changed this. Instead of relying on the JS call stack, React now has its own Fiber nodes. Whenever you update something, React creates a separate work-in-progress Fiber tree in the background (a detached copy of the current one). After all the work is done, React commits it, and the Fiber root starts pointing to the newly committed tree. That’s basically how modern React avoids blocking the UI and handles priorities. People still say React “Virtual DOM” because it’s easy and they don’t understand Fiber architecture or how React really works internally after React 16 btw latest version is (19.x). Pretty interesting right ? #ReactArchitecture #DeepDive #TechExplained #ReactFiber #ReactReconciliation

3

1 Comment

PART 5 ------ Dangers of agentic system: Hallucinations (as I mentioned in Part 3) are huge problem with LLMs and it can never be solved deterministically. And LLMs are the backbone of "agentic" system ! Well the vendors will say agents plan, pursue goal autonomously, blah blah blah. But the issue is when the fundamental perception/reasoning engine is hallucinating, everything built on top becomes garbage in, garbage out at an exponential scale. Hallucinated World Model (state itself is hallucinated so wrong objectives will be pursued and so it cascades below) Hallucinated Goal Understanding   Hallucinated Planning Hallucinated Actions Hallucinated Learning (learning wrong things!) Oh they say, we will put humans in loop then where is the automation. I do understand that in many cases we are overstaffed or some tasks are so simple that it is screaming for automation. THINK AUTOMATION and NOT misplaced view that AI can replace human intelligence. It cannot period!  It can do lots of easy mechanical things in every industry including software.  But, there is no intelligence in reality. It is just patterns at best. Human can factor it into their decision making. Again streamline, downsize, etc. based on automation, innovation, etc. and not on the view that AI is compulsory to achieve the same. Agentic systems, because of hallucination, come with existential risk compared to passive risk from AI query tools. Think about it: What if your infrastructure system is rebalanced causing service level degradation. What if an trading system detects opportunity and executes huge erroneous trade.  Go read about Nx Build fiasco (https://lnkd.in/gsVVnxqa) When the LLM goes down. Your agentic system will not function! What will happen to your business? LLMs require huge energy. What if their is a power crisis. Will business shut down? And have you ever questioned why you require agentic solution. For example, people would use LLMs to create their resume. What will an automated Agentic HR solution do in filtering good candidates. Every one has A+ resume becaue of LLM. Concluding thoughts and few more links: https://lnkd.in/grEVH4Eg

2