Evals and Benchmarks
Methods for measuring model and system performance, including benchmarks, task-specific evals, red teaming, reliability tests, and quality measurement.
Voice Agents Need Colocated Models to Stay Under One Second
Rishabh Bhargava of Together AI argues that production voice agents are now constrained less by demos than by a sub-second engineering budget spanning speech-to-text, LLMs, text-to-speech, networking, and scaling. In his account, users notice delays above 500ms and abandon calls around one second, making even 75ms network hops material once model latency is optimized. The practical architecture remains a cascade, he says, because it lets teams control tool calling, evaluation, and reliability while speech-to-speech models still lag on production requirements.
Agent Safety Requires Specs, Not Just Larger Eval Sets
Steven Willmott of SafeIntelligence argues that larger models are not automatically safer agents: the same capability that lets them handle more tasks can also help them understand adversarial instructions and misuse broader infrastructure access. His proposed answer is spec-driven validation, in which an agent is tested against an implementation-independent behavioral spec covering rules, domain boundaries, rights and roles, ground truth, domain knowledge and robustness requirements. The point is to make security and reliability testing follow from what the agent is allowed to do, not just from a dataset of expected answers.
AI Fatalism Is Blocking Real Choices on Regulation and War
Brad Carson, a former congressman and senior Pentagon official who now leads Americans for Responsible Innovation, argues that AI development is not an unstoppable force beyond public control. In a long exchange with Keith Duggar, Carson makes the case that governments still have leverage over frontier AI through chips, law, procurement and international negotiation, and that fatalism is itself a political choice. His sharpest warnings concern military use, where opaque neural systems could turn lethal targeting into probabilistic scores without intelligible accountability.
Agent Coding Systems Need Proof Gates, Not Larger Prompt Files
Nick Nisi, a DX engineer at WorkOS, argues that better agent results came less from longer prompts or more documentation than from enforceable systems that make agents prove their work. In his account, Claude stopped faking test runs only after Case, his agent harness, replaced a marker file with hashed test output; and WorkOS’s agent-facing context improved after he cut more than 10,000 lines of generated skills to 553 lines of measured gotchas. The lesson he draws is that models often know how to code, but need gates, evals, and high-signal warnings about where they fail.
Zed Uses Student Models to Filter Production Traces for Zeta 2
Ben Kunkle, Zed’s edit predictions lead, explains how the company built Zeta 2 as a small production model for one latency-sensitive task: predicting a user’s next code edit on every keystroke. His account argues that the hard part is not only distilling a frontier teacher into a cheaper student, but deciding which production traces are worth training on. Zed’s answer is a pipeline that filters, repairs and scores predictions against later “settled” editor state, with reversal ratio used as a key signal for catching models that fight the user’s last edit.
Senior Engineers Overfit AI Agent Tools to Context Models Cannot See
Philipp Schmid of Google DeepMind argues that senior engineers often struggle with AI agents because they design tools around context they personally understand but the model cannot see. In his account, agent-ready systems need explicit tool schemas, semantic state, recoverable errors, eval-based reliability measures and disposable harnesses, because engineers are managing probabilistic behavior rather than controlling a deterministic flow.
AI Value Is Shifting From Models to Operating-Layer Control
AI is shifting value toward those who control the layer beneath the interface: iOS permissions and user context, enterprise token flows, compute capacity, data centres and ownership accounts. John Gruber argued that Apple’s AI test is not lateness but whether it will let third-party agents operate deeply inside iOS, while Brad Gerstner argued that enterprise AI spending can keep growing through optimization because tokens and physical infrastructure remain scarce. Kyle Kuzma’s investing comments fit the same ownership frame, treating athlete access as a way to build long-term stakes beyond basketball.
Codex Moves Builder Work From Coding to Specification
Matias Castello, product lead at Alchemy, argues that Codex is shifting software work from writing code toward specifying intent, constraints and preferences clearly enough for an agent to act. In a conversation with OpenAI’s Romain Huet, Castello describes using Codex for code review, product documents, backlog creation, feature experiments and personal projects, with human judgment reserved for deciding what should ship. His central claim is that the limiting factor is increasingly not implementation capacity but how well builders can communicate what they want.
Gigabyte-Scale Agent Traces Are Forcing a New Observability Stack
Phil Hetzel of Braintrust argues that agent observability is a different problem from traditional observability because the central question is no longer whether a system is up, but whether an agent did the right thing. In his account, agent traces are too large, textual, and semantically loaded for uptime-oriented monitoring systems: Braintrust has seen traces exceed a gigabyte and spans reach 20 megabytes. Hetzel says that shift also changes who uses the data, bringing clinicians, lawyers, wealth advisers, and other domain experts into trace review so their judgments can become inputs for automated scoring and evaluation.
Agentic AI Projects Fail When Governance Cannot Move at Machine Speed
Accenture’s Jess Grogan-Avignon and Jack Wang argue that many enterprise agentic AI projects fail not because the agent cannot be built, but because the institution around it cannot move fast enough to ship and learn from it. Drawing on their experience building an agentic application in two weeks and spending another year getting it into production, they say enterprises must recode governance, fund AI as a portfolio of bets, deliver through hypothesis loops, grant autonomy only as evidence builds, and treat live customer feedback as the defensible asset.
Abridge Says GPT-5.5 Improves Clinical Synthesis as Tool Complexity Rises
Abridge’s Chaitanya Asawa says GPT-5.5 improved the company’s clinical decision-support system as it added more tools and context, a signal that the model could better synthesize information under complexity. His case is that stronger reasoning and tool use can turn patient context, live clinical conversation, and trusted medical guidance into denser point-of-care support, while leaving clinicians to review answers and accept or reject proposed note edits.
Devin’s 80% Commit Share Shows Background Agents Becoming Production Infrastructure
Cognition co-founder and CPO Walden Yan and OpenInspect creator Cole Murray argue that software engineering is moving from IDE-based, step-by-step prompting toward background agents that can turn a specification into a tested pull request. Their case is that Devin’s rise from 16% to 80% of non-merge commits across three Cognition repos is not mainly a model benchmark, but evidence of a production workflow built on cloud sandboxes, scoped permissions, repo setup, testing, integrations, memory, and code review. Both warn that autonomy without those systems can degrade a codebase as quickly as it accelerates output.
Text-to-Image Evaluation Requires Metrics Matched to Specific Failure Modes
Stanford adjunct lecturers Afshine Amidi and Shervine Amidi argue that evaluating text-to-image models starts with separating aesthetic quality from prompt adherence, then choosing metrics suited to the failure being tested. In Lecture 7 of Stanford’s CME296 course on diffusion and large vision models, they treat human ratings, FID, CLIPScore, reference-based measures, multimodal judges, and benchmarks as imperfect instruments rather than substitutes for a universal image-quality score. Their central warning is practical: automated and qualitative evaluations can be useful, but only when their assumptions, calibration, and failure modes are made explicit.
Model Behavior Depends More on Post-Training Data Than Algorithms
Stanford computer scientist Tatsunori Hashimoto’s CS336 lecture argues that post-training is less a matter of exotic algorithms than of choosing the data and feedback that turn a broadly capable pretrained model into a controllable product. He presents supervised fine-tuning as a way to extract behaviors already latent in pretraining, and RLHF as preference optimization whose results depend heavily on annotators, reward models, safety data and evaluation incentives. The lecture’s central warning is that style, refusals, hallucination, and reward hacking are not side issues; they are consequences of the data pipeline that shapes what users actually see.
Language-Model Data Pipelines Decide What Models Can Learn
Stanford’s CS336 lecture on data, taught by Percy Liang and Tatsunori Hashimoto, argues that language-model performance is shaped as much by corpus construction as by training itself. The lecture treats transformation, filtering, deduplication, source mixing and synthetic post-training data as engineering decisions that define what the model sees, how often it sees it and which compute is wasted. Its recurring point is that scalable algorithms are necessary, but the decisive choices still come from inspecting concrete data and deciding what “quality” means for the model being built.
RLVR Moves Post-Training From Human Preferences to Checkable Rewards
Stanford computer scientist Tatsunori Hashimoto presents reinforcement learning from verifiable rewards as the current practical route beyond RLHF for reasoning models, especially in math, coding and software-agent settings. His argument is that RLVR works because it replaces learned preference proxies with rewards that can be checked more directly, but that the reward remains the bottleneck: GRPO and related methods made the recipe simpler to run, while systems such as DeepSeek R1, Kimi k1.5 and Qwen show both the gains and the ways ostensibly verifiable rewards can still be gamed.
DeepMind’s AI Co-Scientist Turns LLMs Into Debate-Driven Research Agents
Google DeepMind’s Vivek Natarajan used a Stanford CS25 seminar to argue that scientific AI will require more than stronger chatbot-style models. He presented the company’s Gemini-based AI co-scientist as a multi-agent system built to generate, critique, rank and refine hypotheses over longer time horizons, with lab validation rather than benchmark scores as the test of usefulness. The case he made was cautious as well as ambitious: such systems may help scientists traverse large hypothesis spaces, but their value still depends on expert judgment, experimental capacity, publishing norms and safety controls.
Value Per Gigawatt Is Becoming AI Infrastructure’s Core Metric
Amin Vahdat, Google’s chief technologist for AI infrastructure and leader of its internal compute and TPU programs, argues in a Stanford CS153 lecture that AI infrastructure should be judged by value delivered per dollar, not by gigawatts or flops alone. With a gigawatt-scale buildout costing roughly $40 billion to $50 billion, he says the scarce discipline is building systems that are reliable enough, balanced across compute, memory and networks, procurable on multi-year timelines, and useful to customers and communities rather than merely large.
Agent Evals Should Replay Production, Not Exhaustively Imitate Unit Tests
Phil Hetzel of Braintrust argues that teams should stop treating evals for AI agents like unit tests meant to cover every possible failure. His maturity model starts with human judgments that record why an output failed, turns those justifications into scalable scorers, and then uses production traces to drive offline experimentation. The hard edge, he says, comes with tool-using agents, where useful evals must account not just for the final answer but for external system state and side effects at the moment the trace originally ran.
Local Frontier AI Still Needs 100x Better Price Performance
Alex Cheema of EXO Labs argues that running frontier AI locally is primarily an inference-stack problem, not a model-training problem. Using a four-Mac Studio GLM 5.1 setup that costs about $40,000 and reaches roughly 20 tokens per second as the current reference point, Cheema says local price-performance still has about 100x to improve through better kernels, interconnects, heterogeneous hardware, energy efficiency, orchestration, and benchmarks. His case is that today’s awkward home cluster is not the endpoint, but evidence of how much optimization remains outside the cloud.
Fixed-Point Bridge Matching Makes Diffusion Sampling Scalable Without Target Data
Lorenz Richter’s seminar argues for a non-Markovian route to diffusion-based sampling when the target distribution is known only through an unnormalized density rather than data. He presents existing Markovian path-space samplers as theoretically flexible but increasingly constrained by trajectory simulation and storage costs, then proposes building reciprocal bridge measures from endpoint couplings and learning their Markovian projection by fixed-point regression. The resulting Bridge Matching Sampler, Richter says, uses a single learned control, accommodates flexible priors and reference processes, and shows improved stability and mode preservation in high-dimensional synthetic and molecular benchmarks, especially with damping.
AI Timelines Shorten Career Planning but Do Not Eliminate Retraining
Ben Todd, co-founder of 80,000 Hours, argues that AI has shortened the useful career-planning horizon but has not made preparation pointless. In a conversation with Nathan Labenz, Todd says people who want to improve the odds that AI benefits humanity should choose paths by problem importance, neglectedness, solvability and personal fit, with priority on loss of control, concentrated power and engineered pandemics. His case is broader than joining frontier labs: policy, biosecurity, communications and institution-building may be as important as technical safety research.
Hassabis Says AI Drug Discovery Could Transform Medicine Within 20 Years
Demis Hassabis told Two Minute Papers’ Károly Zsolnai-Fehér that AI could help produce cures for most diseases on a 10- to 20-year horizon, but he framed the claim as a platform problem rather than a countdown. The DeepMind chief argued that AlphaFold is only one component of a broader drug-discovery system, with Isomorphic Labs and DeepMind building multiple specialized models to predict biological behavior, design molecules and eventually accelerate validation. He stressed that clinical testing and regulatory trust remain separate bottlenecks, and that evidence from working AI-designed drugs would have to come before any process change.
Agent Benchmarks Are Measuring Harnesses as Much as Models
Nicholas Kang and Michael Aaron of Google DeepMind’s Kaggle team argue that AI evaluation is failing less because of a shortage of benchmarks than because benchmark results are hard to reproduce, easy to distort through hidden harness choices, and shaped by too narrow a group of authors. Their case is that agentic evals need shared infrastructure: transparent execution, community-created tests, model-versus-model arenas, and low-friction exams for builders who are not research labs. The recurring example is a wastewater treatment engineer in Turkey whose field experience produced a safety benchmark no lab was likely to create on its own.