79 stories
Azeem Azhar's Exponential View #571 covers DeepSeek V4's cost advantage over GPT-5.4, Ukraine's seven-day drone warfare iteration cycle, and new Ember data showing solar output beginning to overtake nuclear on a rolling 12-month basis.
Ethan Mollick describes an MBA experiment where students built startup prototypes in four days using AI agents, then develops a framework for when to delegate to AI based on human baseline time, probability of success, and evaluation overhead.
Ethan Mollick's eighth AI guide since ChatGPT's launch restructures the question of which AI to use around three concepts — models, apps, and harnesses — reflecting what he describes as a fundamental shift away from chatbot-centric access.
Ethan Mollick reviews benchmark progress across multiple AI evaluations, describes a three-person team at StrongDM running a fully AI-automated software factory, and argues the field has entered a phase of unpredictable rolling disruption.
Ethan Mollick argues that chatbot interfaces impose cognitive costs that reduce productivity gains from AI, reviews Anthropic's Claude Cowork with the Dispatch mobile feature, and outlines emerging interface patterns from multiple AI labs.
One Useful Thing's Ethan Mollick describes his early-access experience with GPT-5.5, comparing it to prior models on coding, image generation, and agentic tasks, and notes that the jagged frontier of AI capability has not fully closed.
The Algorithmic Bridge's Alberto Romero surveys over 30 studies from institutions including MIT, Wharton, Harvard, and Stanford on what AI chatbot use does to brain activity, critical thinking, learning, and emotional life, and identifies a common pattern across the literature.
Alberto Romero of The Algorithmic Bridge examines Anthropic's decision to benchmark Mythos Preview alongside Opus 4.7 while withholding the model from public release, and argues it signals a shift in how frontier labs relate to consumer users.
Alberto Romero published an essay at The Algorithmic Bridge arguing that AI-generated content will become a liability as AI agents mediate online discovery. The full essay is behind a paid subscriber paywall.
Alberto Romero of The Algorithmic Bridge contends that writers trying to avoid AI writing signals are bending their prose to an inverse pattern that is itself a form of machine-imposed constraint.
Alberto Romero's weekly digest covers funding news around SpaceX, Cursor, and Mistral; a Dwarkesh podcast with Jensen Huang; and commentary on GPT-5.5 and ChatGPT Images 2.0.
Arvind Narayanan and Sayash Kapoor announce their AI Snake Oil newsletter has been renamed to Normal Tech, signalling a shift from near-term AI hype debunking toward a broader framework for understanding AI's societal effects.
Arvind Narayanan of AI Snake Oil / Normal Tech contends that Moravec's paradox is selection bias dressed as a theory, and that its evolutionary justification rests on claims about neuroscience that AI researchers lack the background to support.
A paper by Arvind Narayanan, Sayash Kapoor, and Justin Curl, published in the Lawfare Research Paper Series, identifies unauthorised practice of law regulations, the adversarial structure of legal proceedings, and mandatory human involvement as bottlenecks that will prevent AI capability gains from automatically translating into cheaper legal outcomes for consumers.
A paper by Stephan Rabanser, Sayash Kapoor, and Arvind Narayanan evaluates 14 models from OpenAI, Google, and Anthropic across 12 reliability dimensions and finds that while task accuracy has improved substantially over 18 months, measures of consistency, robustness, and predictability have advanced only modestly.
A paper by Sayash Kapoor, Arvind Narayanan, and 17 co-authors defines open-world evaluations as a class of long, real-world AI tests distinct from benchmarks, and introduces CRUX, a 17-researcher collaboration that ran a first experiment in which an AI agent successfully submitted a working iOS app to the Apple App Store.
Interconnects author Nathan Lambert argues that the post-Mythos backlash against open-weight AI conflates general capability risks with narrow cybersecurity concerns, and that compute costs and harness complexity place meaningful limits on who can deploy such models at scale.
Interconnects author Nathan Lambert contends that economic pressures are narrowing the field of organisations willing to release fully open, near-frontier models, and that a multi-company funding consortium will eventually be required to sustain them.
Interconnects author Nathan Lambert summarises four parallel projects: an updated ATOM Report on the open language model ecosystem, a completed RLHF book entering production, a free lecture series on post-training, and two co-authored research papers on multi-turn dialogue and meta-reinforcement learning.
In a post on Interconnects, Lambert lists his current beliefs about open-weight model trajectories, arguing that Chinese open-weight labs will face funding difficulties later in 2026, that closed labs have durable advantages in robustness and real-world usefulness, and that economic rather than technical factors will ultimately determine the balance.
In a post on Interconnects, Lambert argues that the Artificial Analysis Intelligence Index and similar composite benchmarks obscure shifting capability domains, benchmark saturation, and the distinction between benchmark performance and real-world usefulness — factors that complicate straightforward claims about how open and closed models compare.
AI #165 from Don't Worry About the Vase catalogues the week of Claude Opus 4.7's launch, noting mixed user reception, a model welfare post that Mowshowitz describes as the most important of the three Opus 4.7 pieces, and OpenAI's release of ImageGen 2.0.
Zvi Mowshowitz's April 2025 monthly roundup in his Substack 'Don't Worry About the Vase' covers platform fraud detection at Lyft and Amazon, open-plan office critique, judicial incentive design, and AI-related operational security concerns at frontier labs.
The Oxford Internet Institute and Department of Education researcher, whose work examines how digital technologies and AI can produce injustices in learning, has been elected to the Academy of Social Sciences Fellowship.
A review of 83 studies on unpaid care and digital technology finds benefits including remote coordination and community support, alongside systematic risks: data privacy violations, carer burnout from always-on connectivity, and inequalities tied to the digital divide.
A survey of 1,198 musicians across Brazil, Chile, the Netherlands, Nigeria, and South Korea finds broad agreement that streaming is essential but pays too little — while revealing sharp country-level differences in how artists experience the platform era.
Several researchers and DPhil students from the Oxford Internet Institute are presenting at ICLR 2026, April 23–27, with papers covering LLM simulation of human behavior, failure prediction from model activations, knowledge distillation, self-explanation faithfulness, and a new reasoning benchmark.
OII Research Associate Johanna Ballesteros argues that Europe's digital sovereignty agenda conflates two distinct goals — securing existing technological dependencies and building capacity for future ones — and that the gap between them explains why the EU remains heavily reliant on US cloud infrastructure.
A recap of the Vector Institute's Foundation Models for Science Workshop in Toronto describes three research challenges: the annotation bottleneck in astronomy, the trust gap in pharmaceutical AI, and the lab-to-production leap, with case studies from astrophysics and drug toxicity prediction.
The Vector Institute held its inaugural Demo Day on January 20 in Toronto, bringing 12 startups from its FastLane program before four judges and nearly 30 venture capital firms, as part of a broader effort to accelerate AI commercialization without taking equity stakes.
Researchers from the Vector Institute, GEMINI, Unity Health, and Diabetes Action Canada developed CRISPNAM-FG, a neural additive model that predicts diabetic foot complications while exposing individual feature contributions, validated on over 107,000 Ontario patients.
The Vector Institute's third annual Remarkable conference featured a 60-poster session on February 20 covering health care AI, safety research, and foundation model architecture, drawing more than 1,500 researchers and industry participants in person and online.
Researchers affiliated with the Vector Institute contributed 48 accepted papers and led 6 workshops at the International Conference on Learning Representations in Rio de Janeiro, covering reinforcement learning, generative AI, autonomous agents, and clinical applications.
Constantinos Daskalakis, an MIT professor and CSAIL principal investigator, received the 2018 ACM Grace Murray Hopper Award for proving that finding Nash equilibria is computationally as hard as finding Brouwer fixed points.
Researchers at MIT's Computer Science and Artificial Intelligence Laboratory have built RoboRaise, a system that uses electromyography sensors on a user's arm to let a robot mirror their lifting motions and respond to nonverbal hand gestures, without speech or programming.
MIT CSAIL and Novartis reported in 2019 on a collaboration testing a Wi-Fi-like device that uses machine learning to measure patient mobility, gait, breathing, heart rate, and sleep stages without wearable sensors, in a trial conducted at Novartis facilities.
A team from MIT's Computer Science and Artificial Intelligence Laboratory published ChainQueen, a GPU-accelerated differentiable physics simulator for soft robots that provides gradient-based feedback on design improvements, addressing a longstanding gap in simulation tools for deformable materials.
A joint MIT and IBM team published GANpaint Studio in 2019, a system that edits photographic images by manipulating units inside a generative adversarial network — and that rejects user commands to place objects in physically implausible locations.
The Allen Institute for AI released WildDet3D, a model that predicts 3D bounding boxes from a single RGB image across more than 13,000 object categories, alongside a dataset of over one million images with 3.7 million verified 3D annotations and training code.
The Allen Institute for AI describes two benchmarks — ScienceWorld and DiscoveryWorld — built to measure whether AI agents can perform scientific investigations rather than recall scientific facts, with top frontier models still completing only about 20 percent of DiscoveryWorld's harder tasks.
The Allen Institute for AI published Branch-Adapt-Route (BAR), a three-stage recipe that trains domain experts independently and merges them into a single mixture-of-experts model, allowing individual skills to be upgraded or replaced without retraining the full model.
METR has published a detailed review of the statistical assumptions underlying its AI time horizon metric, finding that most reasonable alternative modelling choices reduce recent models' 50% time horizon estimates while often increasing 80% time horizon estimates.
METR analyzed 77 records from the NanoGPT speedrun spanning May 2024 to March 2026, classifying each by optimization depth and idea provenance; four recent records were credited to AI agents.
Ben Thompson published a paywalled piece on April 26, 2026 with the headline 'John Ternus and Apple's Hardware-Defined Future, SpaceXAI and Cursor'. The full argument is behind a Stratechery Plus subscription and cannot be verified without access.
Databricks released a reference guide on operational databases that argues traditional OLTP systems lack native support for AI workloads and positions its Lakebase product as a replacement architecture.
Article held pending direct WIRED access, xAI right-of-reply, and independent corroboration of misconduct allegations per STYLE §15.
In Import AI 451, Jack Clark covers Andy Hall's framework for 'political superintelligence' — AI systems designed to help citizens, representatives, and institutions engage with politics more effectively, structured across three layers: information, representation, and governance.
Simon Willison highlighted a Nilay Patel essay from The Verge, calling it 'a superb piece of commentary,' in which Patel argues that the 'software brain' worldview leaves technologists unable to account for why most people do not want automation.
Latent Space's AI News digest ran a headline reporting a DeepSeek V4 Pro and V4 Flash release. The digest URL returned 404 at preparation time, making all specific claims — including the existence of DeepSeek V4 — unverifiable. Auto-published, awaiting editor review.
Simon Willison's blog quoted a tweet by OpenAI's Romain Huet in which Huet said OpenAI merged its Codex coding line into the main model at GPT-5.4 and will not release a separate GPT-5.5-Codex.
Sustainability scientist Amy Luers and optimization researcher Ishai Menache discuss datacenter emissions data, IEA projections, and specific AI applications in energy efficiency and materials science in a Microsoft Research podcast episode.
Shopify CTO Mikhail Parakhin describes three internal AI systems — Tangle for reproducible ML workflows, Tangent for automated optimization loops, and SimGym for customer simulation — and explains why review and CI/CD are now the bottleneck, not code generation.
Stratechery's annual pre-Next interview with Google Cloud CEO Thomas Kurian covers what has actually changed in agents since last year, how Google positions its integration advantage, and where Gemini stands against the competition.
A Latent Space and Unsupervised Learning crossover recorded just after AIE Europe covers the AI coding wars, open model momentum, infrastructure instability, and why memory may be the next major bottleneck.
Claude Code's subscription tiers run from $20 to $200 a month, with usage limits developers say they exhaust in sessions. Goose, an open-source alternative from Block, runs locally with no subscription fees or rate limits.
Simon Willison's llm CLI tool releases version 0.31 with GPT-5.5 support, a new verbosity parameter for GPT-5+ models, configurable image detail levels, and async registration for custom OpenAI model configs.
DeepSeek has released a preview of V4 in two variants, with pricing below closed-source alternatives, a 1-million-token context window, and optimization for Huawei's Ascend chips — the first DeepSeek model targeted at domestic Chinese hardware.
Samsung's mobile chief has warned leadership the company could post its first net loss on smartphones despite strong Galaxy S26 sales, as DRAM and NAND prices roughly double, Ars Technica reports citing South Korean outlet Money Today.
At AIE Miami, a recurring conversation among CTOs and founders centered not on whether to use more AI, but on how to avoid the waste that comes with undirected token volume — depth over breadth in agent workflows.
A NeurIPS 2025 paper from Berkeley introduces a framework that evaluates and optimizes imaging hardware using information content alone, matching end-to-end learned methods while requiring less memory and no task decoder.
A paper from Berkeley AI Research proves that word2vec learns in discrete rank-incrementing steps under mild approximations, with the final representations given by the top eigenvectors of a matrix defined by corpus co-occurrence statistics — equivalent, the authors say, to running PCA.
A sponsored article in MIT Technology Review, produced in partnership with SAP, argues that enterprise AI systems fail not on model performance but on missing business context, and that a data fabric architecture is the proposed fix. SAP's president and chief product officer of Data and Analytics is the primary source.
Writing in Wired, Steven Levy profiles John Ternus — Apple's incoming CEO and 25-year hardware engineering veteran — and argues the leadership transition leaves unresolved how Apple will establish itself in the AI era.
The San Francisco cloud platform—2 million developers, 30 employees, zero marketing spend—abandoned Google Cloud in 2024 to control its own hardware and now claims sub-second deployments and pricing that undercuts hyperscalers by roughly 50 percent, VentureBeat reports.
Berkeley researchers introduce two algorithms that reframe LLM interaction discovery as a sparse recovery problem, enabling feature, data, and mechanistic interpretability at scales where prior methods degrade.
The next generation of large language models will be cheaper, better at long tasks, and structured very differently. MIT Technology Review maps the architectural shifts already underway.
Ted Young told GrafanaCon that reaching CNCF graduation requires stabilizing the long tail of instrumentation packages — a challenge so large that AI coding techniques may be necessary to handle it.
Berkeley researchers introduce GRASP, a gradient-based planner that addresses exploding gradients, non-greedy local minima, and vision-model brittleness in long-horizon trajectory optimization.
The fifth annual New Future of Work report from Microsoft Research documents generative AI's entry into workplaces — faster than previous technologies, with measurable productivity gains for some, declining entry-level hiring, and significant gaps by gender, income level, and language.
A Berkeley researcher proposes a third RL paradigm beyond TD and Monte Carlo that exploits the triangle inequality in goal-conditioned settings, avoiding the error accumulation that limits Q-learning on long-horizon tasks.
Ben Thompson's Stratechery analysis of Tim Cook's 15-year tenure argues Cook was an exceptional scaling CEO who inherited the iPhone at peak growth potential, and frames the question of what Apple's next phase requires.
A paper by Ji-il Park, Inwook Shim, and Chong Hui Kim maps AI technologies to each stage of military Course of Action planning and proposes a system architecture, working from publicly available doctrine because operational systems are classified.
Researchers behind the New Future of Work Report 2025 discuss the intentionality required to build a future where people flourish, the gap between efficiency and the work future worth wanting, and why AI's role as tool versus collaborator is not a semantic question.
An internal memo confirmed by The Register reveals Microsoft is offering voluntary buyouts to US employees at senior director level or below, with eligibility tied to age plus tenure reaching 70 — potentially covering around 9,000 workers.
The Wired podcast discusses Tim Cook's exit and John Ternus's succession, a reported $60 billion SpaceX option on Cursor, and Palantir's 22-point manifesto posted to X.
OpenAI's prompting guide for GPT-5.5 recommends starting from scratch rather than porting existing prompts, and introduces a Codex skill for automated migration with light prompt rewriting.
Wired outlines five failure modes for AI chatbots used in financial contexts: unresolved hallucination, documented sycophancy, privacy risks from data sharing, absence of legal accountability, and research suggesting human advisors disengage when clients also use AI.
TechCrunch's Equity podcast covers Tim Cook's planned September departure, SpaceX's reported $60 billion option on Cursor with a $10 billion breakup fee, and a $5 billion Amazon investment in Anthropic.
Stratechery's weekly digest for the week ending April 24, 2026 covers Tim Cook announcing he will move to Executive Chairman in September, a SpaceX partnership with Cursor that includes an option to acquire at $60 billion, and a dense Sharp China episode on the Strait of Hormuz and semiconductor legislation.