Adrian Cockcroft

Google just released Gemma 4. Open weights. Apache 2.0. Runs on your own hardware. The narrative is compelling: no API costs, no lock-in, full control. AI moving from renting intelligence to owning it. The narrative is also incomplete. The weights are inert. Correct. Gemma 4 cannot phone home. But the moment it does anything useful, connecting to a database, calling an API, processing a file, accessing an external service, the attack surface does not shrink. It expands. There is no cloud provider safety layer anymore. The model is local. The governance is not. Our YIN-SANCTUM becomes more critical with local AI, not less. A locally running agent with database access has nothing between it and the data. The model being on your hardware does not encrypt your queries, prevent unauthorized data access, or verify that the agent is operating within its authorized scope. Our YIN-MCP FIREWALL is the most urgent product in the field right now. Gemma 4 is explicitly designed for agentic use. Every agent connecting to external tools through MCP has zero security on those connections by default. The firewall sits exactly between the local model and the external world. Its value increases when the cloud perimeter disappears. https://lnkd.in/dTb2UDRf Our NeyliaClaw changes this entirely. It wraps Gemma 4, or any local model, with cryptographic governance, verified agent identity, non-memorization enforcement, and a full audit chain. The model runs on your hardware. The governance runs inside every action it takes. Local AI without NeyliaClaw is power without accountability. With it, every decision is proved. https://lnkd.in/e2R3hGPs Gemma 4 running locally with no governance is not a safe private AI. It is an unconstrained agent with access to your files, databases, services, and connections, operating without authorization verification, audit trail, or enforcement layer. The problem was filed. Priority November 23, 2025. 40 USPTO filings. 4300+ claims. #Gemma4 #LocalAI #AIGovernance #AIAgents #MCP #Cybersecurity #OpenSource #AICompliance #Patents #Innovation

1

Tentative observations on Gemini code code generation during a large context problem. I've noticed something that has taken place frequently enough now to hypothesize a pattern and likely explanation. When Gemini is generating code in a shorter session, plus or minus the usual LLM foibles, it tends to work. But. There is some transition point in session length where it will start telling you that it has output code, but no such code was output. All you have is the description of intent and relevance, but not the code. If you point this out, it'll apologize, and try again. And, again, there will be no code. You can even propose alternative document formats like "just render the code to a YAML document and show me that", but no joy. After several repetitions and badgering it, sometimes it tell you that it will just render the code as plain text. This works, although with caveats. Watch out for issues like escaped characters (e.g. for a regex pattern) losing their backslashes. What I think is going on is a confluence of factors: 1. The protections of RoPE for a large context may be breaking down a little. I don't know their RoPE implementation, but possibly as is the case for DeepSeek, it's an approximate algorithm that does not hold up quite as well on large contexts. 2. The normal code output is *not* purely the result of that LLM doing token inference. I'm pretty sure it's a tool integration. There is likely a hidden thinking mode where the code is generated, but then is being routed out of the model for analysis or refinement or formatting (depending on the situation). 3. The tool integration was probably mostly trained on shorter-context problems. If it wasn't fed enough large-context problems, and RoPE gets sketchy, either the tool calls are not being properly generated, or the later response to them is not being properly understood and integrated. End result: Gemini hallucinates having generated anything at all, and once it is in this state, it *stays* in that state except when it completely bypasses attempting to do tool integration (in which case, read carefully for loss of tiny details). Note that this is not the same thing as when you see true context collapse in Gemini on large problems. There the behavior is a bit different. It increasingly loses the ability to remember what its last response was. If it proposed an activity and you respond tersely with "yes" or "sure", it doesn't know what you were agreeing to, and does something else instead. Most often, it re-executes one or two steps back. If you stick around the session for too long at this point, sooner or later it starts spewing utter gibberish that isn't even connected with the rest of the context.

1

2 Comments

I recently joined the stage at DigitalOcean's Deploy conference to talk about a topic close to my heart: the realities of running high-volume inference workloads in production. Thank you so much Vinay Kumar, Paddy Srinivasan Archana Kamath and the entire Digital Ocean team, Hippocratic AI team (Cynara Lilly) for facilitating this. At Hippocratic AI, we’re deploying LLMs across the US for safety-focused, non-diagnostic clinical phone calls. When your system is handling 200-turn conversations, managing P99 and P90 latencies isn't just an infrastructure goal—it defines the entire patient experience. A few of the hard-won lessons we hit on: The Reality of Scale: Why optimization has to move beyond the model layer to the entire orchestration stack (gateways, load balancers, and routing). Open-Source Hackability: Why upstreaming and modifying projects like vLLM and SGLang is essential for survival as architectures shift. Rigor Over Demos: Why deep clinical human evaluations matter more than synthetic benchmarks when safety is paramount. We've published several technical reports on this orchestration work, with more on the way. The full panel discussion here: https://lnkd.in/gSj2Hcfu

66

Getting the AI Hype Cycle Right My recently blog has got some people labeling me as an "AI Skeptic," and I want to clarify: While I agree the current valuations are often 'no bueno,' focusing only on the financial bubble misses the physics of what's happening. Legendary chip architect Jim Keller nailed it: The AI market will crash, correct, and go up again. You have to "low pass filter all that stuff." The Low-Pass Filter Strategy: A filter removes high-frequency noise—the panic, the hype, the frantic 'AI-washing'—and lets the deep, sustained signal pass through. The Signal: As Bill Gates taught us, we always underestimate the change in 10 years. AI’s trajectory to dominate computing is inevitable. The Takeaway: Don't half-ass your strategic thinking by betting on the hype or the crash. Don't miss the revolution by focusing on the jitters. My full analysis on the value equation and filtering the noise: https://lnkd.in/e6K6RW5q #AIStrategy #AIEconomics #DigitalTransformation #Leadership

86

7 Comments

Dwarkesh just dropped a 2.5-hour deep dive with AI pioneer Andrej Karpathy on the Dwarkesh Podcast, and it’s a must-listen, even if it’s a long one. Karpathy is by far has the best and most sober takes on AI progress in the community. Super careful with what he says and how he articulates it and he isn’t too bearish or bullish. Andrej, with his background as former director of AI at Tesla and OpenAI, shared some fascinating insights: 🔹️AGI might be a decade away, but it’s not going to be a dramatic shift—it’ll weave into our lives gradually. 🔹️He dove into the limitations of LLMs, critiqued reinforcement learning, and explained why self-driving cars are taking longer than expected. 🔹️Plus, he thinks AGI could just boost our economy by a steady 2% each year. If you’re into AI, tech, or just curious about the future, grab a coffee, settle in, and give this a listen. It’s worth it! Also on all your usual Podcast platforms.

2

To all the engineers in my network who might or might not know what test-driven development is: The California government’s approach to innovation is like test-driven development. In some cases it actually *IS* test-driven development. I just found out that California’s failed SB 1047 would have “required developers to perform risk assessments of models prior to release, with guidance from the Government Operations Agency.” (wikipedia) There’s one place where the analogy to test-driven development falls short, though. Whereas in TDD you know what you’re testing for, SB 1047 doesn’t even know what it’s trying to regulate, as its goal is to “mitigate the risk of catastrophic harms from AI models so advanced that they are not yet known to exist.” Unfortunately, the state senator who introduced SB 1047 is back at it and still doesn’t realize he is trying to put the cart before the horse. Source: Politico

2

The curious world of micro-apps in the enterprise "aka Enterprise Vibe Coding" I have been thinking a lot about micro-apps in the enterprise. A micro application is a small but important application that a department uses to solve a problem unique to themselves. Some might mistakenly call it shadow it. I can think of two F200 companies I used to work for that had extremely important micro-apps. One company had a spreadsheet that finance used to track IT expenses. If this spreadsheet disappeared or was corrupted, I could see an incredible amount of problems occurring. Another company rebuilt a massively used IOS app under someone's desk via Jenkins. This app had millions of users per day. We knew if this Jenkins/Xcode server went down...we'd have a very bad day. Today, I want to empower the enterprise to 'vibe code'. I want to work with departments to help them vibe code on a platform with an AI assistant and a real no code tool. I want to help them do it in a safe, secure, and thoughtful way with guardrails that are not obvious or intrusive. This is what I always thought DevOps would be. I thought about this problem because we are working on an AI factory. I've been telling Sr. Executives about micro apps, but I'm not sure it was understood. This weekend, I wanted to help executives and peers play around and understand a long form chat I had with Claude w/o my geekiness (how many people can talk about why size matters or the difference between quantization levels?). The result was this Enterprise AI Model Optimizer https://lnkd.in/eEbzM_TP We have the tools for our enterprise customers to solve personal problems today. That spreadsheet could be a react app. That CI build server could leverage templated prompts to deploy the app. Both could be in a controlled environment if we provide the tools. This really changes how I will manage and view shadow IT. I encourage everyone to 'vibe code'

17

5 Comments

This recent memo from the administration shows they are thinking big about how science and technology fuel the nation’s next leap forward. AI, biotech, energy, and space are front and center, but what stands out most is the focus on people and infrastructure: building the workforce, strengthening labs, and connecting agencies to move faster together. It’s a smart reminder that successful innovation requires more than creativity and inspiration. It's needs coordination and integration across a complex ecosystem of moving parts. I’m encouraged to see a call for that kind of alignment, because when agencies share tools, talent, and trust, breakthroughs happen faster. ✨ https://lnkd.in/eEbXfBWg

1

Software is shifting.. From Interfaces you Operate to Systems that Produce Outcomes.. I really enjoyed this piece from Michael Grinich: https://lnkd.in/gHwvkWuJ he articulated in lay-mans terms the history of how we have interacted with computers and what's coming next. We now use chat and detailed context to describe to the computer what you want and it figures out what to do. What's most important is the input. This hockey stick will continue to accelerate as we connect more and more of these Agents to the tools we have today; our computers, our browsers, file systems.. and yes our entire soup to nuts; concept-design-estimation-sourcing-procurement-execution-completion and maintenance of heavy civil construction projects.. Bulk Exchange #sourcingagent #constructiontech #readymix #EPDs

14

Jensen Huang dropped a number at #GTC 2026 that stopped me mid-scroll. "I believe computing demand has increased by 1 million times over the last two years." Jensen Huang, GTC 2026 Not over a decade. Not someday. Two years. Then came the slide that reframed everything for me. The entire AI infrastructure thesis in one frame: ▸ AI Factories are the Industrial Infrastructure of the AI Era ▸ Inference is the Workload ▸ Tokens are the New Commodity ▸ Compute is Revenue And the line Jensen said out loud: "It's now a factory to generate tokens." on the All-In podcast at GTC, Jensen laid out a thought experiment. If a $500K/year developer spent less than $250K on tokens by year end, he'd be "deeply alarmed." My read on that: it's not a pricing benchmark. It's a provocation. It's Jensen saying we are still dramatically underestimating how much value a token can unlock. Here's where I land on all of this. The cost of generating tokens is dropping fast. The value of what those tokens can do is rising exponentially. The gap between those two curves? That's where the real opportunity lives. But and this is the part most people skip the gap only gets captured if the infrastructure beneath it is built right. #Tokens don't generate themselves. #Inference doesn't optimize itself. And enterprises won't run mission-critical agentic workloads on infrastructure that can't guarantee latency, throughput and cost-per-token at scale. Whoever builds that layer , reliably, efficiently and at enterprise grade #owns the #AIstack. The inference delivery platform. At NeevCloud®, this is exactly what we're building toward on our AI SuperCloud platform. Because Jensen is right the AI factory era is here. The question is who builds the pipes that make the factory run. What's your take is token cost the most underloved problem in enterprise AI right now? Drop your thoughts below 👇 Note: I didn't attend GTC in person. Insights are from the official NVIDIA GTC keynote on YouTube and public coverage. Screenshot credit: NVIDIA GTC 2026 official livestream. #NvidiaGTC #AIInference #TokenEconomy #AgenticAI #EnterpriseAI #AISuperCloud #NeevCloud

34

Anthropic just doubled Claude Code limits after locking in a massive compute deal with SpaceX. Claude Code's 5-hour usage window is now 2x for Pro, Max, Team, and Enterprise subscribers. Peak hour limit reductions are gone for Pro and Max users. API rate limits on Claude Opus got a significant bump too. The compute comes from SpaceX's Colossus 1 data center in Memphis, a facility powered by over 220,000 Nvidia GPUs. Anthropic is getting access to more than 300 megawatts of capacity through the deal. This is Anthropic's answer to Claude usage exploding faster than their existing infrastructure could handle. SpaceX built Colossus 1 primarily to power Grok and xAI, so Anthropic is renting capacity from a rival's hardware. When even competing labs share infrastructure to keep up with demand, you know AI usage is scaling at a pace nobody fully planned for. What does this mean for solo devs and small teams shipping on Claude Code? -- source: https://lnkd.in/eXFwSw6m

1

Today the Pulse team published a deep dive on why a single “accuracy” score doesn’t tell you if a document extraction system will survive in production. The goal here is to lay out an introductory but still rigorous evaluation methodology - we have an exciting open-source benchmark building on this research coming out very soon. Let’s do the math: take 1,000 pages, each with 200 data elements. A model that’s 98% “accurate” on paper still produces 4,000 incorrect values. Now make some of those: 1/ Broken reading order that scrambles multi-column layouts 2/ Tables with shifted columns or missing headers 3/ Cross-page context lost entirely That’s enough to silently corrupt an entire dataset without throwing a single error. We’ve processed hundreds of millions of pages and built a multi-axis evaluation framework to measure what actually matters: reading order validation, region-level ANLS, reading order accuracy, TEDS for table structure, and continuity checks across page boundaries. The result? Fewer silent data corruptions, more predictable performance, and pipelines that keep working on the next million documents you haven’t seen yet. Full technical write up in the comments!

36

2 Comments

AI investment aimed at improving developer productivity is one of the hottest conversations in engineering leadership right now. Bill Buckley, a member of VPE Forum, shared a perspective I think is especially insightful. https://lnkd.in/eFiFsJDB

12

I’ve tried and used many tools surrounding LLMs, such as LangChain/Graph, LlamaIndex, vector stores like Qdrant, and full AI platforms like Azure AI Foundry. I don’t understand why none of them offer RAG applied to their own documentation - it seems like such an easy win. Many of these technologies are new and evolving quickly, so getting grounded information from places like ChatGPT can be difficult. Trawling through documentation feels unnecessary in the age of LLMs. Maybe someone can enlighten me.

18

9 Comments

🚀 The Opus Post-Mortem: Decoding Claude 4.7 🧠 Is your AI feeling a bit sluggish lately? 🐢 If you noticed Claude 4.6 acting differently, you weren't imagining things. Our latest deep dive into the "Opus Post-Mortem" reveals the mechanics behind the phantom degradation of 4.6 and the massive leap forward with the 4.7 release. 📉➡️📈 The telemetry data for Opus 4.6 showed a significant 73% decrease in output character length, indicating a thinking collapse. This led to a critical 33.7% file reading failure rate and a 12x increase in required user interruptions. 📉❌ But here is the good news: Opus 4.7 is here to fix it all. 🛠️✨ Key highlights of the 4.7 upgrade: • Adaptive Default Thinking: Smarter allocation of reasoning power. 🧠✅ • X-High Effort Level: Exclusive high-performance mode for complex tasks. 💪🚀 • Biomolecular Reasoning: A staggering >2x jump in performance. 🧬🔥 • Production Tasks: A 3x multiplier metric for real-world reliability. 🏭🔝 While 4.7 introduces a 1.3x tokenizer tax due to upgraded text processing, the results in live environment tests are undeniable. From finance chart analysis to complex SaaS financial modeling, 4.7 consistently catches errors its predecessor missed and provides actionable, specific insights. 📊🏦 The 4.7 release isn't just a fix; it's a structural upgrade that restores functionality and returns deep thinking to your workflow. 🌟🙌 Full Video here: https://lnkd.in/dNVdHjUd Have you made the switch to 4.7 yet? Share your experience with the new reasoning capabilities below! 👇 ♻️ Repost to help your network understand why their AI just got a major brain upgrade. ➕ Follow [Deven Goratela](https://lnkd.in/dVt7VtDu) for the latest insights into staying ahead in the rapidly evolving world of AI and automation. #ClaudeAI #Anthropic #ClaudeOpus #GenerativeAI #PracticalAI #AIGrowth #TechInsights #Automation

1

🤓 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

AI-assisted coding isn’t about fewer keystrokes, but greater clarity, deeper thought, and scalable precision. Diwank Singh’s "Field Notes From Shipping Real Code With Claude" guides developers through battle-tested AI workflows: 🍃 Playground Mode for swift exploration. 📐 Pair Programming Mode using CLAUDE.md to shape architecture and uphold style. 🚦 Production Mode fortified with human judgment, stringent tests, and anchor comments. This is developers orchestrating AI to amplify craftsmanship. If you're exploring how AI can elevate your coding practice, start here. Read more → https://lnkd.in/gG68gjxr #AICode #EngineeringBestPractices #DevOps

8

3 Comments

I recently gave a talk at the inaugural Forum on Emerging Technologies (https://lnkd.in/dfizSGPY) at UC Berkeley's Jacobs Institute called "The Agency of Yogurt." Five minutes to argue that the human meaning-space of any ordinary product category — yogurt, crackers, insurance — is a fog of intersecting habits, beliefs, identities, constraints, and emotions. Which may seem counterintuitive — yogurt is a straightforward proposition, no? Not really. Any category you can name, from the self-evidently complex like a laptop to the seemingly simple like dish soap, is layered with meaning and ritual and aspiration and routine. To fully grasp yogurt, to contend with the roles it plays for people, is archeological. Too alive to flatten into a 2×2 grid, too densely, richly complex to contain on any whiteboard. For the latter half of the 20th century, the prevailing move was to compress that fog. Segment. Validate. Eliminate ambiguity. Help the observer feel oriented, gain the (often as not falsely) reassuring sense that they had the thing pinned down. The new move is to structure the fog without flattening it. Navigate it. Play in it. The difference between looking at a paper map and stepping inside an immersive virtual environment. That's what we built. SemField takes any product or category and computationally maps its meaning-space. You choose a grammar, and thousands of intersections open up — most nonsensical, some genuinely surprising. Then you play. A cognitive assist without overdetermining anything in advance. https://semfield.ai/ — first 20 to sign up get $5 in free tokens. Use code YOGURT at checkout. Yes, really. More on the underlying thinking: https://lnkd.in/dbaceh8Y Horst Rittel taught at Berkeley. His insight: design problems are wicked because understanding the problem and solving it are the same activity. He knew that in 1967. We just didn't have instruments that could hold the complexity while you worked. The quietly strange thing about a tool like SemField is that it doesn't automate a phase of the work. It dissolves the boundaries between phases entirely: the researching is the thinking is the prototyping is the strategy. That's unsettling — in an exhilarating way, true — but it's a real departure from the mode of working all of us have known all our working lives. It unsettles us, and we made the thing.

58

5 Comments