The Hidden Complexity of the Grid Most software developers view spreadsheets as simple, flat grids. We assume an LLM that can write complex Python scripts will easily navigate a financial model. It turns out that spreadsheets are deceptively difficult for AI. When a human opens Microsoft Excel, they instantly grasp the visual layout. They spot the revenue tables, highlight assumptions, and track the flow of numbers. An AI agent sees none of this. To a model, a spreadsheet is a chaotic maze of coordinate pairs, hidden dependencies, and ambiguous labels. Nuno Campos and his team at Witan Labs spent four months teaching coding agents to interact with spreadsheets. They discovered that success requires moving beyond basic tool calls and building high-fidelity feedback loops. The REPL Breakthrough In early designs, the team built fifteen separate tools for the agent. The agent had to make sequential tool calls to read cells, write formulas, and check values. This approach created massive lag. High latency led to frequent timeouts. ```javascript // Instead of 15 separate tool calls, agents run a single script const data = sheet.getRange("A1:B10"); const total = data.filter(r => r[0] === "Revenue").reduce((sum, r) => sum + r[1], 0); ``` The breakthrough came when they replaced these fifteen tools with a single Node.js REPL. This interface allowed the agent to write a complete script to execute multiple actions in one round trip. Crucially, they chose a REPL with persistent state over a simple run-and-die code mode. With persistent state, the agent defines variables in one turn, analyzes the result, and uses those same variables in the next step. This design shortened the scripts the agent wrote, allowing more room for step-by-step reasoning. The Need for High-Fidelity Feedback An agent cannot write good code without a compiler, and it cannot build a spreadsheet without a calculation engine. To enable true self-correction, Witan Labs built two core validation engines: a formula engine to calculate cell values and a rendering engine to show the sheet as an image. They quickly learned a vital lesson about engine quality. If your custom calculation engine only supports half of Excel's formulas, you make the agent worse. The agent writes a valid formula, but the incomplete engine throws an error. The agent then gets confused, assumes its logic was wrong, and tries to fix code that was already correct. High-fidelity engines are non-negotiable for the validation loop. How to Measure True Progress When optimizing their system from 50% to 92% accuracy, the team had to overhaul their evaluation methods. Relying solely on an LLM-as-a-judge proved fragile. Changes in the evaluator's prompt would shift scores, making it impossible to tell if the agent had actually improved. They pivoted to deterministic testing. They took a golden spreadsheet with known inputs and outputs, fed the same inputs to the agent's generated sheet, and checked if the outputs matched. This black-box verification provided a rock-solid metric. Many apparent reasoning failures were actually simple framework bugs. By analyzing traces and fixing their plumbing, they cleared the path to production-grade reliability.
LLM
Products
Mar 2026 • 2 videos
Lighter month. Rees and 20VC with Harry Stebbings covered LLM across 2 videos.
May 2026 • 3 videos
High activity month for LLM. AI Coding Daily, TechCrunch, and AI Engineer among the most active voices, with 3 videos across 3 sources.
Jun 2026 • 3 videos
High activity month for LLM. AI Engineer among the most active voices, with 3 videos across 1 sources.
Jul 2026 • 1 videos
Minimal activity. LLM mentioned in 1 videos from 1 sources.
AI Coding Daily notes LLM's persistent memory, as discussed in "I Tried Deleting CLAUDE.md. Here's What Happened," while Laravel Daily highlights their consistency in generating Laravel code, and The Riding Unicorns Podcast addresses competition in the LLM space (3 mentions).
- Jul 8, 2026
- Jun 8, 2026
- Jun 6, 2026
- Jun 1, 2026
- May 12, 2026
The Monolithic Memory Problem in Enterprise AI Enterprise AI has hit a wall. Despite the explosion of LLM capabilities, from prompt engineering to deep agents like Replit, Jira tickets aren't moving faster. Raj Navakoti, a staff software engineer at IKEA, identifies a core structural issue: institutional knowledge is a monolith. He compares the current state of AI agents to the protagonist of the film Memento, who possesses high skill but cannot hold memory for more than 15 minutes. This creates a perpetual state of disorientation where agents excel at general tasks but fail at specific, domain-heavy requirements. The industry currently pushes a retrieval-heavy strategy involving RAG and MCP servers. However, plugging an MCP server into a broken, monolithic knowledge base is like connecting a high-speed pipe to a dry well. Navakoti argues that roughly 40% of critical organizational knowledge is "tribal"—it lives in people's heads and never hits Confluence or Slack. Another 20% is outdated, and 20% is unreliable. When agents fail, they aren't just failing to process data; they are exposing the structural gaps in how a company documents its existence. Switching from Knowledge Push to Demand-Driven Pull Traditional knowledge management relies on a "push" strategy: engineers attempt to document everything upfront and push it toward the agent. This is inherently inefficient. Navakoti proposes a "pull" strategy inspired by how we onboard new human employees. We don't ask a junior dev to memorize the entire company wiki before their first commit; we give them a task, and they pull the necessary information as they hit roadblocks. This demand-driven context approach turns the agent from a passive consumer into an active knowledge manager. In this framework, the agent is intentionally given a problem it will likely fail to solve. This failure is the catalyst. Instead of giving up, the agent generates a checklist of exactly what it doesn't understand. It identifies missing API definitions, unclear business logic, and undocumented system behaviors. This "demand extraction" surfaces the tribal knowledge that documentation efforts usually miss because they don't know what they don't know. The Agent Lifecycle and the Curation Loop The methodology operates in a cycle similar to Test-Driven Development (TDD). In TDD, we write a failing test first to define the requirement. In demand-driven context, we provide a "failed problem." The cycle moves through four distinct phases: 1. **Problem Assignment**: An agent is tasked with a real-world issue, such as a root cause analysis for a production incident. 2. **Discovery of Gaps**: The agent attempts retrieval via RAG but identifies specific missing entities or outdated information, assigning confidence scores to its current context. 3. **Human Interjection**: A domain expert fills the specific gaps identified by the agent's checklist. This is surgical documentation rather than a broad, exhaustive effort. 4. **Curation and Storage**: The agent takes the new information, curates it into a structured format (like Markdown), and saves it to a persistent repository. By running this cycle across multiple incidents, the agent's confidence score improves measurably. Navakoti demonstrated that over 14 incident cycles, an agent's confidence in handling tasks jumped from 1.5 to 4.4 out of 5. The agent builds its own "cache" of curated context blocks that are significantly more useful than the raw monolith of Confluence pages. Why GitHub is the Ultimate Knowledge Repository A controversial but practical element of Navakoti's framework is the storage layer. While many look to expensive SaaS solutions for knowledge management, he advocates for storing curated context in GitHub repositories. The reasoning is purely engineering-driven: GitHub provides built-in version control, Pull Request reviews, and conflict resolution. When multiple agents and human experts contribute to a shared knowledge base, data conflicts are inevitable. Treating knowledge as code allows teams to use the same rigorous CI/CD pipelines for their documentation as they do for their software. If an agent proposes a documentation update based on a resolved incident, a human expert can review that PR to ensure the logic is sound. Once merged, that knowledge is permanently indexed and available for all other agents in the ecosystem. Navigating the Domain with Meta Models Beyond raw text, the framework benefits from a Meta Model—a map of how different domain entities relate to one another. An agent needs to understand that a notification service failure affects specific business processes and relies on certain APIs. Without this map, the agent is just guessing which files to retrieve. The Meta Model acts as a navigation layer. It allows the agent to reason about the "blast radius" of a change or the dependencies of a specific system. When the file structure of the knowledge base reflects this Meta Model, retrieval becomes deterministic rather than probabilistic. Instead of hoping the vector database finds the right chunk, the agent knows exactly which branch of the knowledge tree to explore. Scalability and the Future of Agentic Operations Critics of this approach often point to the human cost of answering agent questions. Navakoti acknowledges this but argues the cost is front-loaded. You are effectively performing a one-time "denial of service" on your engineers to fix a decade of documentation debt. Once the 20% of most-used knowledge is curated into context blocks, the agent becomes semi-autonomous. The goal is to move the agent from being a pure consumer of information to a guardian of it. As context windows expand—with Claude now supporting 1 million tokens—the technical limitation is no longer memory size, but memory quality. By using failure as a scanner, enterprises can finally map their "unknown unknowns" and build a knowledge base that actually helps the AI move the Jira board.
May 5, 2026The high stakes of murky information We are currently witnessing the birth of a new information funnel. Every breakthrough in technology brings a period of chaos, and Campbell Brown is sounding the alarm: the large language models currently dominating our lives are essentially "slop" when it comes to high-stakes information. In the pursuit of coding efficiency and mathematical precision, the tech giants have largely ignored the nuanced, murky world of news and geopolitics. This isn't just about a broken link; it's about the erosion of the shared reality required for a functioning society. If we don't fix the funnel, we risk raising a generation that lacks the tools to discern truth from sophisticated hallucination. Moving from engagement to truth The fundamental mistake of the social media era was optimizing for engagement. We learned the hard way at Meta that human beings react most strongly to emotional triggers and opinion validation. My perspective is that Forum AI represents the necessary pivot. We need to move away from "what do people like?" and toward "what is real and truthful?" Enterprise demand will be the catalyst for this change. While a teenager might tolerate a chatbot's creative liberties, a bank making credit decisions or a government agency assessing geopolitical risk cannot. The liability is too high for theater; the market is now demanding actual reliability. Expert reasoning over generalist guesses Scaling trust requires more than just smart generalists or automated box-checking. To build a truly reliable benchmark, you must architect systems that capture the reasoning of elite experts like Tony Blinken or Neil Ferguson. It is about training LLM judges to mirror the nuances of human consensus. We are seeing a massive gap where Google Gemini pulls sources from propaganda sites and ChatGPT lags days behind on breaking news. Fixing this requires a commitment to source selection and the inclusion of missing perspectives, moving beyond the "left-leaning bias" that currently plagues most foundation models. A mandate for AI literacy There is a profound disconnect between the visionary rhetoric of Silicon Valley and the actual experience of the consumer. While leaders talk about curing cancer, the average user is getting wrong answers to basic health questions. We need to implement AI literacy alongside traditional media literacy. This isn't just a challenge for students; it’s a requirement for the teachers and the professionals who are currently being told that their jobs are on the line. We must bridge the gap between the "hopefulness" of the tech elite and the "low levels of trust" in the general public. The opportunity of the neutral model Despite the controversy surrounding political mandates, the underlying principles of truth-seeking and neutrality are the only path forward. We have a rare opportunity to use AI to push back against the echo chambers and filter bubbles that have defined the last decade. If we optimize for truthfulness rather than clicks, we can reconstruct a consensus reality. The power to decide these principles is the ultimate leverage in the modern economy. Those who build the most truthful systems won't just win the market—they will secure the future of informed discourse.
May 1, 2026The Industry Icon Meets the Living Legend: Larry Hryb Joins Commodore There is something poetic about Larry Hryb, the man better known to millions as Major Nelson, joining forces with the brand that arguably started it all for the home computing generation. For over twenty years, Larry Hryb stood as the bridge between Microsoft and the Xbox community, pioneering the very idea of a corporate personality who actually talks *to* people rather than *at* them. Now, he’s taking that veteran expertise to Commodore to serve as a community development advisor. Perry Fractic, the current president and CEO of Commodore, has been aggressively rebooting the brand, and bringing in a heavyweight like Major Nelson is a massive statement of intent. This isn't just about nostalgia; it’s about navigating the tricky waters of a modern tech relaunch. Commodore has recently faced its share of legal drama—mostly involving other entities claiming rights to the name—but the successful release of their new Commodore 64 hardware project shows there is still a massive appetite for that signature breadbox aesthetic. Larry Hryb understands how to build a global ecosystem, and if anyone can help Commodore introduce its legacy to a new generation of creators and enthusiasts, it is the man who helped make the Xbox 360 a household name. Pushing Silhouettes: Sweet Fighting Plus Two on the Spectrum When we talk about hardware limits, we usually discuss 4K textures or ray tracing. But the real magic happens when you try to cram a game designed for high-end arcade boards into 128K of memory. Sweet Fighting Plus Two is a brilliant homebrew project that brings the Street Fighter II experience to the ZX Spectrum 128K. This isn't just a gimmick; it’s a masterclass in compromise and creative engineering. Developed by the team at ZX Press, the game features a staggering roster of 12 legendary fighters. To keep the gameplay fluid and the frame rate playable, the developers made a bold stylistic choice: they kept the health bars and UI in the Spectrum's limited color palette but rendered the actual fighters and backgrounds in high-contrast black and white. This eliminates the "color clash" that usually plagues the system and allows for incredibly detailed sprites that actually look like their arcade counterparts. Playing this on original hardware—or via an interface like the DivMMC—is a reminder that "impossible" is just a challenge for the right developer. It’s a love letter to the 128K hardware that proves these old machines still have plenty of fight left in them. The Audio Magic of Amiga OutRun and the Vinyl Revival Reassembler has been doing the Lord’s work for the Amiga community over the last year. His port of OutRun to the Amiga is legendary because it didn't just try to mimic the arcade; it used the original 68,000 assembly code to ensure the physics and timing were pixel-perfect. But the real standout of that project was the music. Reassembler meticulously converted the iconic Sega soundtrack into Amiga tracker formats, giving it that unique four-channel, 8-bit "mod" flavor that only Paula (the Amiga sound chip) can provide. That soundtrack is now getting a physical release on a stunning translucent red and blue starburst vinyl, embedded with glitter. This is DJ-friendly, mastered loud at 45 RPM, and features original hand-painted artwork by Sam Miller. It’s a fascinating intersection of retro gaming and high-end audio collectibles. For Amiga fans, this is more than just a record; it's a trophy representing the community's ability to take 30-year-old hardware and outshine the official ports of the 1980s. Whether you're blasting "Magical Sound Shower" in a Ferrari Testarossa or spinning it on a modern turntable, the craftsmanship here is undeniable. Victory Against the Trolls: The Liberation of Cookie's Bustle One of the most significant wins for game preservation recently comes from the Video Game History Foundation and their battle for a Japanese adventure game called Cookie's Bustle. Released in 1999 by a developer called Rodic, the game is a genre-defying masterpiece that involves an anthropomorphic bear, international sports, and intergalactic war. For years, however, talking about or sharing footage of this game was impossible because of a persistent copyright troll named Brandon White (operating as Graceware SL). This individual used DMCA takedowns to wipe almost all mention of Cookie's Bustle from the internet, despite having no legitimate claim to the IP. The Video Game History Foundation, in collaboration with legal counsel, finally stepped in to expose the lack of evidence for White's ownership. This victory resulted in Yuki (the trade association) suspending takedowns for the title. It’s a critical case study in how "orphan works" are vulnerable to exploitation and why organizations that fight for the public domain are so vital to our digital heritage. Now, researchers and fans can finally document and translate this bizarre, wonderful piece of history without fear of legal retribution. The Intelligence Divide: AI and the Lutris Controversy No technical discussion in 2026 is complete without the elephant in the room: AI. Lutris, the beloved open-source game manager that makes playing Windows games on Linux a breeze, has found itself at the center of a heated debate. The lead developer of Lutris admitted to using Claude, an LLM from Anthropic, to help write code and catch up on maintenance during a personal health crisis. This sparked a firestorm on GitHub, with some users labeling the new commits as "slop." The developer defended the choice, arguing that AI is an augmentation tool, not a replacement, and pointed out that Anthropic had recently push-backed against government contracts. However, critics like Liam Dawe from GamingOnLinux argue that AI companies are sucking up hardware resources and infringing on copyrights, making their use unethical in the open-source world. The developer has since removed the "co-authored by Claude" tags to avoid further drama, but the controversy highlights a growing rift in the DIY tech community: do we embrace these tools to save projects from burnout, or do we reject them to protect human craftsmanship and hardware availability? Demons in the Dashboard: Doom on Home Assistant Finally, we have to talk about the man who brought Doom to Home Assistant. Developer Frank Nijhof took the ultimate "can it run it?" challenge and integrated the 1993 classic directly into his smart home dashboard. This isn't just a simple web wrapper; it’s a deep integration that treats Doom as a smart device. When you start "ripping and tearing," Home Assistant knows. You can set up automations so your office lights turn a hellish red the moment the Doom binary sensor flips to "on." It tracks player stats, session duration, and even displays a daily "Doom Fact" on your wall tablet. It’s completely unnecessary, technically brilliant, and perfectly captures the DIY spirit of pushing software into places it was never meant to go. Whether you're playing on a PC, a Mac, or your kitchen's smart hub, the fact that we are still finding new ways to experience id Software's masterpiece is a testament to the enduring power of great design.
Mar 13, 2026The public markets are currently treating the software sector like a terminal patient, but Eran Zinman isn't interested in the funeral rites. As the co-CEO of monday.com, Zinman has watched his company’s valuation compress even as fundamentals remain resilient. The disconnect between business operation and market sentiment has birthed a series of doomsday prophecies: that AI will allow everyone to build their own software, that foundation models will swallow the application layer, and that agents will render interaction platforms obsolete. Zinman dismisses the noise, arguing that we are entering the most aggressive growth phase in the history of technology. Death of the seat-based economy The most structural threat to the legacy SaaS model isn't just the existence of AI, but the collapse of the headcount-linked pricing model. For twenty-five years, software value was tethered to the number of human beings clicking buttons. If AI can perform 80% of the work previously done by humans, the traditional per-seat license becomes a liability for the vendor and a resentment for the customer. monday.com is currently navigating a pivot toward consumption-based pricing, acknowledging that value must be tied to output rather than payroll size. This shift is radical. It requires a total re-engineering of the go-to-market strategy, the product interface, and the revenue recognition models that investors use to judge health. Critics argue that moving away from seats will cannibalize revenue, but this perspective ignores the massive expansion of the Total Addressable Market. Zinman contends that while headcount spend might decrease, software spend as a percentage of corporate budgets will skyrocket. Companies that currently spend 8% of their budget on software and 70% on humans will see those ratios invert. The opportunity isn't about protecting the existing $1.3 billion in revenue; it’s about capturing a piece of a market that is set to expand by two orders of magnitude as software moves from being a tool for tracking work to a tool for doing the work. Vibe coding and the illusion of simplicity The concept of "vibe coding"—the idea that non-technical users can simply describe a software requirement to an AI and have it perfectly manifest—has become a viral existential threat. When a journalist built a functional clone of monday.com in a few hours using AI, it sent a shockwave through the investor community. Zinman views this as a fundamental misunderstanding of what makes enterprise software valuable. There is a massive delta between generating a user interface and maintaining a scalable, collaborative, and secure infrastructure that works across a ten-thousand-person organization. Building the first 10% of a tool is easy; maintaining the remaining 90% through years of organizational change is where the value lies. While consumer-grade apps might be vulnerable to this democratization of development, enterprise environments demand a level of cohesion that fragmented, self-coded tools cannot provide. monday.com is positioning itself not as a tool that can be replaced by a vibe-coded script, but as the underlying operating system where those agents and scripts are orchestrated. The goal is to move from being a system of record to a system of action, where the complexity is managed in the background while the user focuses on the strategy. Why the model companies won't kill the apps A persistent fear in the VC world is that OpenAI, Anthropic, and Google will move up the stack and render application companies like Salesforce or monday.com irrelevant. History suggests otherwise. Zinman points to the early days of AWS, when skeptics predicted Amazon would capture all enterprise value because they owned the infrastructure. Instead, the ease of infrastructure created a boom in application development. The model providers are focused on the massive opportunity of being the "backbone" of intelligence. Selling, implementing, and supporting complex enterprise software requires a completely different DNA—a sales-heavy, handheld process that model companies are ill-equipped to execute at scale. Furthermore, intelligence without context is useless. An LLM is brilliant but blind to the specific, undocumented strategies and workflows that live within a company's walls. The application layer provides that context. monday.com sees its future as the bridge between raw intelligence and the specific context of a business. By being the horizontal platform where humans and agents collaborate, they capture the data that makes the AI effective. The model providers might provide the engine, but the application layer provides the fuel and the steering wheel. Playing offense in a defensive market While most SaaS companies are cutting headcount and hunkerng down to survive the "SaaS Apocalypse," monday.com is maintaining a mid-teens headcount growth. This decision seems paradoxical to some, but Zinman views it as an offensive necessity. You cannot capture a 100x TAM expansion by playing defense. The company is aggressively integrating AI into its own internal operations—replacing its 100-person SDR team with agents and automating its customer support—not to reduce the total number of employees, but to reallocate human talent toward the high-leverage tasks of building the next generation of the product. Internal morale during a 60% stock drawdown is a management hurdle, but Zinman uses the low valuation as a psychological reset. When the market prices your company at a level that implies the business is worth nearly zero after accounting for cash, the only response is to go "all in." This involves taking big, calculated risks on vertical offerings like CRM and Service, and betting the entire platform on an agentic future. The companies that emerge from this cycle as winners will be those that didn't just survive the transition, but used the chaos to rewrite the rules of their industry. For monday.com, the objective is to move past the era of being a work management tool and become the essential orchestration layer for a world where agents do the majority of the heavy lifting.
Mar 2, 2026