Daniel J. R.

When Anthropic introduced the Model Context Protocol, they promised to simplify using agents. MCP enables an AI to understand which tools rest at its disposal : web search, file editing, & email drafting for example. Ten months later, we analyzed 200 MCP tools to understand which categories developers actually use. Three usage patterns have emerged from the data : Development infrastructure tools dominate with 54% of all sessions despite being just half the available servers. Terminal access, code generation, & infrastructure access are the most popular. While coding, engineers benefit from the ability to push to GitHub, run code in a terminal, & spin up databases. These tools streamline workflows & reduce context switching. Information retrieval captures 28% of sessions with fewer tools, showing high efficiency. Web search, knowledge bases, & document retrieval are key players. These systems are likely used more in production, on behalf on users, than during development. Everything else including entertainment, personal management, content creation, splits the remaining 18%. Movie recommenders, task managers, & Formula 1 schedules fill specific niches. MCP adoption is still early. Not all AIs support MCP. Of those that do, Claude, Claude Code, Cursor top the list (alliteration in AI). Developer focused products & early technical adopters are the majority of users. But as consumer use of AI tools grows & MCP support broadens, we should expect to see a much greater diversity of tool use.

84

13 Comments

We live in an alphabet soup of AI benchmarks - MMLU, GPQA Diamond, HLE, LiveCodeBench, SciCode, AIME 2025, IFBench, AA-LCR, Terminal-Bench Hard. Benchmarks are supposed to measure progress. Instead, they’ve become the AI equivalent of EBITDA: technically informative, practically an exercise in costume design. Two predictable failures: ▪️Saturation: Everyone hits 90% eventually. Once frontier models ace MMLU or GSM-8K, the test stops meaning anything. It’s like grading Olympic sprinters on whether they can jog a 10-minute mile. Impressive the first time, meaningless once everyone’s jogging across the finish line. The scores go up, the signal goes down. So researchers invent the next “harder” exam, and the cycle repeats. ▪️Optimization: Benchmarks get hacked. Models fine-tune for the test, not the world. You can wire a model to dominate HumanEval without making it a better programmer. You can goose GSM-8K with chain-of-thought tricks that would never help in an actual math class. In some cases (hello, Llama 4), vendors allegedly submitted “special” benchmark-only models. That’s not progress - it’s leaderboard cosplay. If you’re buying AI solutions, this is where things get tricky. On paper, everyone looks great. But none of that tells you whether Vendor X’s chatbot will resolve your customer tickets, or Vendor Y’s model will flag fraud in your transaction flow better than the rest. Benchmarks test the generic. Your business runs on the specific. Which is why you need a golden dataset: a private exam with your data, your edge cases, your definition of success. Fraud logs in finance. De-identified patient notes in healthcare. Customer chat transcripts with the slang and sarcasm only your users know. This is boring work, but it’s also the only thing that matters. Because: 1. Different vendors optimize differently. One for reasoning, one for retrieval, one for latency. All of them might be “top-ranked” broadly - but only your dataset will reveal which matches your needs. 2. The real world is messy. Benchmarks (especially static ones) sanitize the world. Your data has noise, ambiguity, system interactions, edge cases, domain jargon. A model might get 95 % on MMLU but tank on your rare edge case. 3. Contamination happens. Tests leak into training data, so models memorize. Private datasets cut through that. 4. The tails matter. In public benchmarks, everyone clusters near the top. The spread shows up in the long-tail failures. Benchmarks are like standardized tests. They’re useful for broad comparisons, but if you’re hiring, you don’t just want to know if a candidate can ace the SAT. You want to know if they can actually do the job you’re hiring for. AI is no different. So yes, clap politely when a model sets a new record. But when it comes to procurement, bring your own exam. The only benchmark that matters is the one no one else has - yours.

124

9 Comments

ICYMI. AI is driving the cost of intelligence to zero. This innovation is going to spark rapid and massive transformations in tough industries: eg drug discovery, advanced mnfr, energy, mobility, defense, etc. https://lnkd.in/gRTinSxp

56

11 Comments

Lara Translate — August 2025 Release Highlights https://laratranslate.com This month, we focused on making the quality feedback smarter, documents more reliable, and supporting Latex and XLIFF 2.0 to empower scientists and localization experts. What’s new: • Lara Quality Feedback: Evaluation now accounts for your translation memories and glossaries, making feedback more reflective of your own style and terminology. • Lara Doxuments: Added support for LaTeX and XLIFF 2.0. Recent improvements (password-protected PDFs, DOCX with tracked changes) helped reduce document processing errors by 17%, bringing success rates to 94.3% over the last 30 days. • Lara Interpreter: Redesigned flow allows language selection directly in the UI and voice input for faster interactions. • Lara for Google Sheets™: Improved evaluation feedback in LARAEVAL() and smoother UX. Some cool features missed the release at last minute, so we may come with a bonus Mid September release.

88

3 Comments

We’re coming to MIT, Harvard, Stanford, Berkeley, and online to tackle one question: how do we build an aligned AI future? Lightning talks from top PhDs at each school + a keynote from our own D. Scott Phoenix. Topics on the table: -- How to align economic incentives with AI alignment, biodefense, interpretability, and more. -- The physical constraints + bottlenecks that will shape our future. -- What becomes exponentially more valuable (and what becomes worthless) on the path to AGI. AI is moving faster than anyone expected. The world urgently needs more researchers building safety solutions, and startups can make safety the default at scale. Link in comments.

67

1 Comment

AI is no longer just trained on data. It’s being raised in environments. Billions are flowing into synthetic training grounds where AI agents practice: – Virtual offices to learn scheduling and negotiation. – Simulated warehouses to master logistics. – Entire artificial marketplaces to test decision-making. This is the shift: from studying the past (datasets) to shaping behavior in invented environments. And here’s the power play: whoever builds these environments sets the rules. Not just what AI knows — but how it acts. That’s the new empire being built. It’s not about bigger models or faster chips anymore. It’s about who controls the environments where AIs learn how to operate — and what hidden incentives and biases get baked in. Are we ready for a handful of companies to design the operating conditions of the machines that will run our future? And who has a role to play in this? #AI #Business #Transformation #Next https://lnkd.in/eSnz6bjy

115

1 Comment

💊 Where are all the AI drugs? 🧐 There’s been a lot of noise lately about AI “revolutionizing” drug discovery. A recent WIRED article by Veronique Greenwood captured the optimism well — suggesting that with tools like large-scale automation and generative models, we’re finally starting to see real traction. The article is an interesting and insightful take, I suggest you read it (link in the comments). However, two cents from me here. The reality is more sober. Despite the hype, very few companies have actually gotten AI-generated drug candidates into human trials. Fewer still into Phase II. In fact, the article only names two: Recursion and Insilico Medicine. ☝And here’s where it gets interesting. One of those programs mentioned, REC-994 from Recursion, was actually an in-licensed molecule that’s since been deprioritized after mixed clinical results, as far as I know. That’s not exactly an AI-born success story IMO. That leaves only Insilico’s ISM001-055, for idiopathic pulmonary fibrosis . Not only is it AI-designed, it’s still active in Phase II, with multiple other candidates moving through the pipeline — from target discovery to clinical testing, all driven by their end-to-end platform 🔥. That’s a critical distinction. There’s a difference between using AI to support R&D and building a platform that can consistently generate novel compounds and push them through real-world validation. So yes, AI is making progress in biotech — but we’re early, and success is still super rare. Which is why what Insilico is doing stands out, in my opinion, not because I am an investor in the company, but because it is the objective reality from what I see on the market at this moment. And by the way, since we touched AI drug discovery, Insilico also quietly rolled out a set of pretty serious upgrades and they explain them in their recent PharmaAI webinar (linking in the comments). There, things like faster physics-based modeling (Alchemistry 2.0), better tools for peptide screening and biologics design, and new features in PandaOmics and DORA that make omics analysis and research writing faster and more usable. They even opened up access to several aging clocks through an open-source package. Kudos to Alex Zhavoronkov and Alex Aliper and the team for pushing the boundaries! 👏

41

12 Comments

Lessons from the Past – Kosmix and Innovative Search Interfaces Long before today’s generative AI revolution, Kosmix was already challenging traditional search interfaces. Founded in the mid-2000s, Kosmix aimed to organize the web around topics rather than keywords, pioneering the concept of "algorithmic topic pages." I had the pleasure of being an early team member of Kosmix. These topic pages were automated, multimedia-rich dashboards tailored to each search query, categorizing and curating information from across the web into coherent, interactive experiences. Kosmix’s categorization engine intelligently identified content types relevant to each query, creating unique, dynamic presentations. One notable innovation was "Tweetbeat," designed for real-time experiences during live events such as the FIFA World Cup. It seamlessly integrated Twitter feeds, filtering real-time posts alongside scores and updates, creating an immersive, interactive search result experience. Kosmix’s groundbreaking approach highlighted critical lessons: the value of context-aware, dynamic content presentations, the need for effective user engagement without overwhelming users, and the complexity of integrating real-time data into search results. Today, generative AI-powered search platforms leverage these insights, employing advanced AI to deliver contextually relevant, dynamic, multimodal search experiences that echo Kosmix's visionary concepts. Kosmix’s story highlights the power—and the complexity—of rethinking the search interface. In the concluding post of this series, we'll envision the future of search, reimagined with generative AI. #SearchInnovation #TechHistory #WebEvolution #UserExperienceDesign #DynamicContent #AIinSearch #InteractiveMedia #InformationRetrieval #RealTimeDataIntegration #FutureOfSearch

31

1 Comment

A model called the Tiny Recursive Model (TRM)—with a minuscule 7M parameters—is outperforming massive foundation models (like Gemini 2.5 Pro and others) on complex, hard puzzle tasks like Sudoku, Maze solving, and the ARC-AGI 1 & 2 benchmarks! Reading this paper made me smile...why? Feels like an embodiment of a real-world "Strange Loop" a la Gödel, Escher, Bach. The network is designed to recursively feed its own output back into its input, creating a self-corrective, self-refining loop until converging on a solution. So is the process of internal self-reflection (the recursive loop) a more crucial ingredient for flexible, generalized intelligence than perhaps sheer scale? Feels like a real score for efficient algorithmic innovation vs. brute force scaling! #AI #MachineLearning #DeepLearning #LLMs #RecursiveReasoning #StrangeLoop #ARCAGI

123

6 Comments

AI is not just hype...it is a layered ecosystem. From Nvidia’s chips, to Anthropic's innovative models and Perplexity’s interface, we are witnessing a reshaping of how tech value is created and captured. I broke this down live on CNBC and followed up with a Medium post that goes deeper into the AI stack: infrastructure, models, and applications and the players leading each layer (Medium post in comments). https://lnkd.in/gaYixDjp

71

6 Comments

Some highlights from this week's episode of the MAD Podcast with Boris Cherny, the creator of Claude Code at Anthropic - a major hit in coding AI: 𝐓𝐡𝐞 "𝐀𝐜𝐜𝐢𝐝𝐞𝐧𝐭𝐚𝐥 𝐑𝐨𝐜𝐤𝐞𝐭𝐬𝐡𝐢𝐩": Claude Code started as a simple internal tool at Anthropic and became an instant hit, leading to its public launch and an incredible $400M in annualized revenue within just five months. 𝐓𝐡𝐞 𝐑𝐢𝐬𝐞 𝐨𝐟 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐂𝐨𝐝𝐢𝐧𝐠: This is a new paradigm. Instead of manually editing code based on AI suggestions, developers now describe what they want to an AI agent, which then handles all the necessary work, becoming a true collaborative partner. 𝐒𝐮𝐛-𝐚𝐠𝐞𝐧𝐭𝐬 𝐟𝐨𝐫 𝐓𝐞𝐚𝐦𝐰𝐨𝐫𝐤: A powerful feature lets you create specialized "sub-agents" for different roles like a project manager or a QA engineer. This allows Claude to divide and conquer complex tasks just like a human team. 𝐀 𝐒𝐡𝐚𝐫𝐞𝐝 𝐌𝐞𝐦𝐨𝐫𝐲 𝐒𝐲𝐬𝐭𝐞𝐦: Claude Code uses a simple system of claude.md files to remember knowledge about a codebase. This allows teams to build a collaborative, shared memory bank that benefits everyone. 𝐌𝐚𝐬𝐬𝐢𝐯𝐞 𝐏𝐫𝐨𝐝𝐮𝐜𝐭𝐢𝐯𝐢𝐭𝐲 𝐁𝐨𝐨𝐬𝐭 𝐚𝐭 𝐀𝐧𝐭𝐡𝐫𝐨𝐩𝐢𝐜: The tool has drastically reduced technical onboarding time at Anthropic from weeks to just a few days. It also makes it easier for non-experts to build things, democratizing creation. There's a lot more - thoughts on the AI coding wars, advice to founders building on top of LLMs, advice to young software developers Watch the full episode here: https://lnkd.in/ePP3mmWg #AI #coding #softwaredevelopment #MachineLearning #ClaudeCode #madpodcast

26

1 Comment

I gave a speech at the largest AI conference in NA - Ai4 - Artificial Intelligence Conferences in Vegas last month, titled "Industry Landscape of Spatial AI“. The full presentation is out and you can watch it below: https://lnkd.in/gvAYQDcJ After spending over a decade in frontier technology as a Silicon Valley VC, I truly believe Spatial AI is one of the most exciting and underrated niches in AI today. A few simple reasons why: 1️⃣ The AI space feels saturated — but Spatial AI is still a blue ocean. 2️⃣ LLMs have fundamental limitations (as Yann LeCun argues) — they don’t natively understand the physical world. 3️⃣ Spatial AI bridges perception with action — it’s where digital intelligence meets real-world geometry. 4️⃣ It underpins robotics, AR/VR, autonomous systems — all trillion-dollar markets waiting to scale. 5️⃣ Advances in sensors, GPUs, and 3D data pipelines make this the right timing. 6️⃣ Most importantly: Spatial AI builds the foundation for embodied intelligence — the step beyond text and images. Special shout out to Michael Gold to collaborate in this journey. We’re just at the beginning. 🚀 #ai4 #spatialAI #spatialcomputing #VC #PE #investment

33

3 Comments

GenAI Startups Ranking – 2025 Generative AI companies are scaling faster than any prior tech wave, with multiple players hitting unicorn status in under two years. This year’s ranking shows rapid growth not only in foundational model providers but also in AI-native tools, enterprise platforms, and vertical applications. While the U.S. leads in volume, China and Europe are emerging as strong competitors. The market is diversifying quickly, and leadership is no longer limited to those building base models.

194

14 Comments

"Dual-use" is easy. The hard part is dual-funding. Everyone obsesses over the tech. Drones, AI, quantum, biotech. But that’s not the bottleneck. The bottleneck is money. Here’s the paradox nobody talks about: 🔷 Civilian VCs love the market size—but panic when the word “defence” appears in the deck. Optics risk > investment return. 🔷 Defence funds? They won’t touch you until you’ve got a product at TRL 6+, certifications done, and a MoD customer lined up. By then, you’re already half-dead. 🔷 Result: Startups with world-class tech die not in the lab, but in the funding gap. Why dual-funding matters: 1️⃣ Two clocks, one startup Civilian markets want growth in 12 months. Defence markets want compliance in 60. You need investors who understand both clocks can’t be reconciled. 2️⃣ Capital is signalling power A €5M cheque isn’t just runway. It’s political proof that your tech is legit. Ministries move faster when Sand Hill Road or Berlin VCs already stamped approval. 3️⃣ Narrative arbitrage The same drone can be pitched as “infrastructure inspection” to civilians and “ISR capability” to defence. Dual-funding means mastering narrative duality. 4️⃣ Grey money vs. clean money Accepting purely defence cash too early can make you toxic to civilian partners. Accepting purely civilian cash can make you invisible to defence primes. You need both, by design. The lesson: Dual-use technology is not the problem. Dual-funding structures are. If Europe wants to win, we don’t just need accelerators. We need investors with dual-thesis funds, sovereign guarantees, and the guts to back companies before regulators and primes feel comfortable. If you’re not designing your cap table as carefully as your product, you won't make it. #DefenceInnovation #DualUse #VentureCapital #Funding #DeepTech #Security #Geopolitics

65

11 Comments