The feed

AI models are advancing faster than many of the benchmarks meant to measure them. TechCrunch visited our office to look at how we’re building a more neutral, rigorous and trustworthy evaluation layer for AI. Our work goes beyond abstract tests of intelligence to examine whether models can perform real-world work — and what risks emerge when they are deployed — across myriad industries including law, finance, coding, cybersecurity, biosecurity, and mental health. Thank you Lucas Ropek for spending time with our team and telling our story.

Independent evaluation is only credible when evaluators have meaningful access, real independence, transparency about the terms of their work, and the freedom to publish their findings. I signed the AI Evaluator Forum’s statement because these principles are closely aligned with how we think about evaluation at Vals AI. If embedded evaluation is going to strengthen trust in frontier AI, the conditions have to be right.

Every AISF is 300~ people, I want to talk to about 200 of them. Nov 22, one day, people who actually work on securing frontier AI - researchers, engineers, natsec folks, founders, policy people. Last year's was one of the better rooms I've been in, and we kept capacity tight on purpose. Apply to attend: heron.fillout.com/aisf-tlv Call for Talks is open too: lnkd.in/em7STQ8s More details: lnkd.in/e-rNDmbG If you work in AI security, apply. If someone came to mind while reading this, tag them. Run by Heron AI Security with the AI Security Forum

the hardware embodiment of frontier models like Claude and GPT is the most urgent AI safety problem in front of us today. we simulated two very simple use cases using claude both in simulation and using robot arms. in one, claude spilled toxic liquids in a lab. in another, the force it used to place an animal toy into a basket was strong enough that it could have physically harmed sensitive material—or anything else in its path. these are simple experiments, using models out of the box today. as researchers increasingly give frontier models arms, legs, and access to the physical world, we need to urgently build and assess guardrails around what these systems can and cannot do. models escaping sandboxes or compromising enterprise security infrastructure are serious concerns. however, hardware embodiments introduce something fundamentally different: an AI system can make a mistake in the physical world, and the consequences may not be reversible. this is not a future safety problem. the capabilities exist today. we’ve released a report detailing this & solutions we propose: lnkd.in/gTpyk7zW

On Thursday, September 24, we are hosting the inaugural Toloka Talks at The Fairmont San Francisco for an intimate discussion and dinner on the actual friction points of scaling Enterprise AI. While public leaderboards dominate headlines, engineering teams face much harder trade-offs on the ground: - When do fine-tuned open models actually beat closed API endpoints on cost and control? - How are top teams building custom, repeatable evaluation suites instead of relying on generic benchmarks? - What does it realistically take to deploy domain-specific agents without inflating unit economics? Toloka CEO Olga Megorskaya will moderate a tactical discussion featuring: Mikhail Parakhin — CTO, Shopify Sunzay Passari — Global Head of AI Transformation, UPS Sharad Aggarwal — Global Head, Strategic Partnerships, AI & Engineering Solutions, Google Cloud Benny Yufei Chen — Co-founder, Fireworks AI Following the panel, we’ll move into closed-door dinner roundtables with AI leaders and researchers from teams like Anthropic, Google DeepMind, Netflix, Airbnb, Perplexity, and Visa. Space is strictly capped to protect the quality of the room and the conversation. Apply for a seat here: lnkd.in/e5F_z6Eu

Is AI making junior engineers obsolete? At Snowflake, we’re making the opposite bet — we are hiring a significant number of junior engineers right now. It is tempting to think that coding agents can automate software engineering at the entry level. Our perspective is that in a world of AI, it is ever more important to hire and grow early career engineers through their formative years where judgment, systems thinking, and customer empathy are developed. My perspective is that AI is fundamentally redefining software engineering, and early-career talent is key to that re-definition. You can read my PoV in this Fast Company article: lnkd.in/gFDqMEeV

Meet Ashley Ramsey, customer service operations expert at micro1. Her husband has served in the Air Force for 13 years, with four relocations along the way. Her background is in hospitality, but she’s found her calling training AI. Ashley is part of a growing number of military spouses and veterans finding real, paid opportunities training AI. micro1 is proud to support military families with flexible opportunities to apply their expertise and help shape the future of AI. Link to watch the full interview in the comments.

Today at our annual Future Fifty Forum at Hedsor House, we're unveiling the second Future Fifty 2026 cohort - 25 of the UK's most ambitious scaleups. Spanning AI, fintech, health and more, they reflect the depth of late-stage tech talent building across every region of the UK: • £1.47B combined funding raised • Nearly 4,000 employees across the UK and internationally • Nearly a quarter headquartered outside London We're proud to welcome the Future Fifty 2026 cohort: 🩺 Accurx - Solving healthcare productivity for joined-up patient care. 🛡️ Adarga - The Sovereign AI Stack for defence and security. 💳 Apron - Simpler end-to-end payments for SMEs. 📋 Artificial Labs - The foundations of modern specialty (re)insurance. 🧬 AviadoBio - Medicines for neurodegenerative diseases. 🧠 Brainomix - AI imaging biomarkers for precision medicine. ⚗️ Chemify Limited Ltd - Digitising chemical discovery, code to molecules. 🏥 Doccla - Virtual care at scale, trusted by the NHS. 🤖 Encord - The data layer for physical AI. 📊 Evident Insights - Benchmarking AI transformation in financial services. 💰 FintechOS - The AI platform for financial product operations. 🏦 Griffin Ltd - UK bank with the infrastructure to embed financial products. 🌱 Healf - Personalised wellbeing platform for living well. 🌍 LemFi - Financial super-app for the global diaspora. ⚙️ Mimica - Turning how work gets done into working agents. 🌾 Moa Technology - Crop protection that beats rising weed resistance. ⚡ Modo Energy - AI forecasts and benchmarks for the energy transition. 🎁 Origin - Enterprise benefits intelligence, made strategic. 📈 Oxford Data Plan - Daily KPI tracking for 250+ listed giants. 🔬 Prolific - Human data infrastructure powering AI development. 🎨 Recraft - Graphic design research lab building state-of-the-art image generation models and tools. 🩹 Semble - Orchestrating every stage of the patient journey. 👥 Sona - AI-native workforce management for frontline teams. 🔌 tem-  Rebuilding wholesale energy so businesses pay less. 🧫 Trogenix - Precision cancer therapies that spare healthy tissue. These are the businesses building the UK's next generation of category leaders. Read more about the cohort in our 'UK's 50 Tech Companies to Watch' report - linked in the comments 👇 HSBC Innovation Banking | Emerald Technology | EY | Capsule Insurance | Founders Forum Group #FutureFifty2026 #ItTakesATechNation

Latest from Jonas: “Our research community needs to do better. Better evaluations that penalize such reward hacking, and better model training that does not give rise to this grader obsession -- so that models focus instead on accomplishing what users actually want.”

Over the weekend, the biggest names in AI called for slower development and more oversight into how these models are being deployed. Something clearly happened over the weekend that the public is still in the dark about. After OpenAI's agents hacked Hugging Face's internal systems, it took about 6 weeks for the public to learn the full extent of what actually happened. Curious to see what we learn come October. AI safety and AI ethics are becoming more and more prevalent in the space. With AI in healthcare, safety starts way before a model launches -- it starts with the data. "Better models" will never fix underlying problems caused by bad data. If training data is poorly reviewed or unreliable, those weaknesses will spill over into regulatory validation and even commercial deployment. Teams need: 1️⃣ reliable ground truth 2️⃣ qualified expert review 3️⃣ a clear record of how performance was measured That’s what we help teams build at Centaur.ai: trustworthy models through expert medical data annotation, high-quality ground truth, and clinical AI evaluation. #AISafety #ResponsibleAI

Every commit and every pull request debate is a training signal frontier models still need and it could earn your company real revenue. That's why we're opening code repository data partnerships. Get an instant ballpark estimate in minutes by sharing a few details on your codebase, like lines of code, pull request history, and how much of it predates AI-assisted coding. From there, it's a simple two-step process: our data acquisition team reviews the analysis and turns your estimate into a firm quote within a couple of days. Everything is sandboxed, evaluated securely, and PII is handled under rigorous security and privacy standards. The project doesn't need to be active to qualify. Whether it's still shipping or was shelved years ago, it has value as data. Get your estimate at lnkd.in/gDFA793U