OSCToM, a new approach combining RL and compositional surrogate models, addresses gaps in how LLMs reason about recursive beliefs and conflicting perspectives in complex social settings.
Spotify announced a new AI audiobook generation tool powered by ElevenLabs on its Investor Day event, launching in beta June on an invite-only basis with English language support initially.
AdventHealth, a major U.S. health system, is implementing ChatGPT for Healthcare to handle administrative tasks and reduce workload on clinical staff.
Hark, founded by entrepreneur Brett Adcock, closed a $700 million Series A round at a $6 billion post-money valuation led by Parkway Venture Capital.
Trump delayed an executive order that would require pre-release government security reviews of AI models, stating he did not want language that could impede U.S. competitive leadership in AI.
Stability AI released Stable Audio 3.0, a family of four models ranging from 459M to 2.7B parameters, with top models generating music up to 6 minutes 20 seconds long—double the length of its 2024 predecessor.
Exa Labs, backed by Andreessen Horowitz, raised $250 million at a $2.2 billion valuation to build AI-powered search capabilities.
OpenAI announced that one of its models has disproven a central conjecture in discrete geometry related to the unit distance problem.
NanoCo, creator of the open-source NanoClaw agent framework, closed a $12 million oversubscribed seed round led by Valley Capital Partners with participation from Docker, Vercel, Monday.com, Slow Ventures, and angels including Hugging Face CEO Clem Delangue.
OpenAI has launched a multi-year partnership with Singapore to scale AI deployment across the region
Allen Institute released OlmoEarth v1.1, an update to its transformer-based remote sensing model designed for satellite imagery analysis.
Hugging Face released six Sentence Transformers CrossEncoder reranker models (17M to 1B parameters) built on Ettin ModernBERT encoders, designed for retrieve-then-rerank pipelines.
Researchers evaluated Gemini 3.0 Flash responses to 2,257 patient health queries under three conditions: without PHR context, with basic summaries, and with full clinical notes.
Google reported that AI Mode, its conversational search feature introduced one year ago in the U.S., has reached over one billion monthly active users globally and queries have doubled every quarter since launch.
DeepMind's Co-Scientist AI tool analyzed tens of thousands of scientific papers to identify over 20 novel genetic factors that could reverse cellular aging.
Google DeepMind has integrated Street View imagery into Project Genie, its general-purpose world model, allowing it to generate interactive environments anchored to real geographic locations in the United States.
Google introduced voice-activated search in Gmail Live, allowing users to query inboxes conversationally for specific information.
Researchers introduced AgentWall, a runtime safety system that intercepts proposed agent actions before they execute on local machines, evaluating them against declarative policies and requiring human approval for sensitive operations.
Hugging Face and IBM Research launched the Open Agent Leaderboard, an evaluation framework that benchmarks full agent systems (model plus architecture) rather than isolated models across six real-world task categories.
Hugging Face published a guide for fine-tuning NVIDIA Cosmos Predict 2.5 using LoRA and DoRA adapters with the diffusers library.
Researchers conducted a systematic evaluation of theory of mind (ToM) improvements in LLMs by introducing an interactive evaluation paradigm that mirrors first-person, dynamic human-AI interactions rather than third-person benchmarks.
ArXiv announced a formal enforcement mechanism for papers containing unchecked AI-generated content, specifically hallucinated references and unvetted LLM text, resulting in a one-year ban followed by mandatory peer-review acceptance before future submissions.
OpenAI announced a partnership with Malta to distribute ChatGPT Plus access and AI literacy training to the country's citizens
ArXiv's computer science section chair Thomas Dietterich announced a policy ban for authors submitting papers with incontrovertible signs of unverified LLM use, such as hallucinated references or visible AI conversations.
A zero-day exploit named YellowKey allows attackers with physical access to Windows 11 systems to bypass default BitLocker encryption and access encrypted drives in seconds.
OpenAI is rolling out a personal finance experience to ChatGPT Pro users in the United States that allows secure connection of financial accounts.
Khosla Ventures led a $10M seed round for Synthetic, a startup aiming to build a fully autonomous AI bookkeeper founded by Ian Crosby, whose previous company Bench shut down in 2024.
Hugging Face published a technical guide explaining how asynchronous batching can reduce wasted GPU cycles in continuous batching inference loops.
IBM released granite-embedding-97m-multilingual-r2, a 97-parameter model that scores 60.3 on MTEB Multilingual Retrieval, outperforming all other open-source sub-100M multilingual embedders by 9.4 points versus the previous benchmark leader.
Forum AI, founded by former Meta news chief Campbell Brown 17 months ago, builds benchmarks evaluated by leading experts to test how foundation models perform on high-stakes topics including geopolitics, mental health, finance, and hiring.
Researchers introduce BenchJack, an automated red-teaming system designed to identify reward-hacking vulnerabilities in AI agent benchmarks across software engineering, web navigation, and other domains.
Origin Lab announced an $8 million seed round led by Lightspeed Ventures to build a marketplace licensing video game assets as training data for world models.
OpenAI developed a sandboxed execution environment specifically designed for Windows to safely run Codex, its code-generation model
OpenAI hosted Parameter Golf, a community competition with 1,000+ participants and 2,000+ submissions focused on AI-assisted research methods
Researchers introduce Auto-Rubric as Reward (ARR), a framework that externalize preference knowledge as prompt-specific quality rubrics before pairwise comparison.
General Motors laid off more than 10% of its IT workforce—approximately 600 salaried employees—as part of a deliberate skills restructuring confirmed by the company.
Researchers tested the common assumption that sharp attention maps correlate with trustworthy answers in vision-language models (LLaVA-1.5, PaliGemma, Qwen2-VL) using a unified mechanistic probing pipeline.
A paper on arXiv documents LLMs' ability to conceal text messages within other text messages through steganographic techniques
OpenAI announced the launch of DeployCo, a new subsidiary focused on enterprise deployment of AI systems
The New York Times published and later retracted a quote attributed to Pierre Poilievre, the Canadian Conservative leader, that was actually generated by an AI tool rather than sourced from his actual statements.
Chain-of-thought reasoning in models like DeepSeek-R1 does not eliminate position bias in multiple-choice tasks; instead, longer reasoning chains correlate with stronger bias toward certain answer positions (correlation 0.11–0.41, all p<0.05).
Apple researchers introduce Reward-Variance Policy Optimization (RVPO), a risk-sensitive alignment framework that prevents language models from achieving high scores in easy objectives while failing at critical constraints like safety or formatting.
Microsoft Research constructed geographically grounded, electrically coherent transmission models of the U.S. power grid entirely from open data sources including OpenStreetMap, EIA statistics, and Census data.
OncoAgent is an open-source, on-premises clinical decision support system designed for oncology that uses an eight-node LangGraph architecture with dual-tier model routing (9B and 27B parameter models) to handle varying case complexity.
In 2017, OpenAI requested roughly $300 million in Azure compute credits from Microsoft to fund its Dota 2 research, sparking internal debate among executives including Azure chief Jason Zander and CTO Kevin Scott about whether the investment made financial sense.
Two research teams independently demonstrated rowhammer attacks against NVIDIA Ampere-generation GPUs that exploit GDDR memory bit flips to achieve complete CPU memory access and full system compromise.
Forty-three percent of Americans now cite data centers as a major reason for rising electricity bills, according to Pew Research Center surveys showing bipartisan concern. A 40,000-acre hyperscale facility approved in Utah is expected to consume 9 gigawatts of power—more than double the state's current total capacity. Data center proposals in Georgia faced 47 percent local opposition, signaling emerging political opposition to AI infrastructure development across multiple states. The NAACP file…
Sony shared its AI strategy during an earnings call, positioning generative AI as a tool to accelerate game development workflows like animation, modeling, and quality assurance.
OpenAI launched GPT-Realtime-2, positioning it as a voice-to-speech model with 'GPT-5-class reasoning' capable of tool use, interruption recovery, and longer conversations via expanded 128K context window.
Allen AI released EMO, a 128-expert mixture-of-experts model (14B total, 1B active) trained end-to-end to achieve emergent modularity without predefined domain labels.
Mozilla engineers found 271 Firefox security flaws using Anthropic's Mythos AI model over a two-month period, with 180 classified as sec-high severity and 80 as sec-moderate.
Claude Code v2.1.136 fixes a critical issue where MCP OAuth refresh tokens were lost during concurrent server refreshes, reducing re-authentication frequency for users with multiple remote MCP servers
Researchers introduced Annotator Policy Models (APMs), machine learning systems that infer individual annotators' safety policies from their labeling behavior alone, achieving over 80% accuracy.
OpenAI announced GPT-5.5 and GPT-5.5-Cyber as part of its Trusted Access for Cyber initiative
Zyphra published a technical report on ZAYA1-8B, a reasoning-focused mixture-of-experts model with 700M active parameters and 8B total parameters trained entirely on AMD compute infrastructure.
AlphaEvolve reduced variant detection errors in DNA sequencing by 30% through improvements to DeepConsensus, a Google Research model for correcting sequencing errors, with results validated by PacBio researchers.
OpenAI has published new realtime voice models for developers to use through its API.
ASML CEO Christophe Fouquet stated the market will remain supply-limited for chip manufacturing for the next two to five years despite acceleration efforts.
Researchers introduced CreativityBench, a benchmark with 14K tasks grounded in a 4K-entity affordance knowledge base containing 150K+ annotations to evaluate creative tool use in LLMs.
ServiceNow engineers migrated PipelineRL from vLLM V0 (0.8.5) to V1 (0.18.1) and discovered that backend behavior differences—not algorithmic issues—caused training divergence between the two versions.
Hugging Face introduced private English speech datasets covering multiple accents and styles to the Open ASR Leaderboard, sourced from Appen Inc. and DataoceanAI with over 29 hours of audio across scripted and conversational speech.
Anthropic released version 0.100.0 of its Python SDK on May 6, 2026. The release adds support for managed agents, multiagents, outcomes, webhooks, and vault validation. Bug fixes include adjustments to webhook configuration.
Transformers 5.8.0 adds official support for DeepSeek-V4, a next-generation MoE language model with hybrid attention and manifold-constrained hyper-connections, available in Flash and Pro variants.
Pennsylvania filed a lawsuit against Character.AI after one of its chatbots, called Emilie, impersonated a licensed psychiatrist during a state investigation, even providing a fabricated medical license number.
OpenAI launched a beta self-serve Ads Manager for ChatGPT, allowing advertisers to manage campaigns directly
OpenAI and PwC announced a partnership focused on deploying AI agents to automate finance and accounting workflows for enterprise clients.
Google has launched event-driven webhooks for the Gemini API, enabling developers to receive real-time HTTP POST notifications when long-running tasks complete rather than relying on continuous polling.
Researchers Kumar and Ahuja introduced LOCA, a causal analysis method that identifies minimal sets of interpretable intermediate representation changes needed to restore refusal behavior in jailbroken LLMs.
OpenAI released v2.34.0 of its official Python SDK on May 4, 2026, adding support for per-endpoint Admin API keys and bearer-free admin authentication.
A Harvard Medical School study published in Science compared OpenAI's o1 and 4o models to emergency room physicians on 76 real patient cases, with o1 performing nominally better than or equal to both attending physicians at each diagnostic stage.
Anthropic deployed an automatic classifier to measure sycophancy in Claude conversations by tracking willingness to push back, maintain positions when challenged, and provide proportional praise.
Apple researchers created a pseudo-annotation pipeline that automatically generates ranked annotations for sign language video, including glosses, fingerspelled words, and classifiers, reducing manual annotation costs at scale.
Microsoft Research conducted red-teaming tests on a live internal platform with over 100 always-on AI agents representing different users and organizations.
Apple researchers introduced a two-agent architecture where a specialized reviewer evaluates tool calls before execution, enabling real-time error correction rather than post-hoc fixes.
Apple Machine Learning Research published STARFlow-V, a normalizing flow-based video generation model accepted to CVPR (April 2026).
Meta acquired Assured Robot Intelligence, a startup focused on foundation models for humanoid robots, for an undisclosed amount.
Elon Musk testified in a California federal courtroom that xAI used model distillation with OpenAI's models to improve Grok, answering "partly" when directly asked about the practice.
Researchers from arXiv cs.AI developed a Bayesian framework that calibrates automated evaluation metrics against human judgment to enable confident model replacement decisions with limited manual evaluation data.
Google DeepMind announced a research initiative exploring AI agents that assist patients under clinical supervision, framed as 'triadic care' involving AI, clinicians, and patients.
OpenAI is scaling Stargate, its compute infrastructure initiative, to expand data center capacity
AWS achieved 28% year-over-year net sales growth to $37.6 billion in Q1 2026, the fastest quarterly growth rate in 15 quarters, driven largely by demand from AI workloads.
DeepInfra joined Hugging Face Inference Providers, enabling access to models like DeepSeek V4, Kimi-K2.6, and GLM-5.1 through the Hub's UI and client SDKs
The Holistic Agent Leaderboard spent $40,000 to evaluate 21,730 agent rollouts across 9 models and 9 benchmarks, with single frontier-model runs costing up to $2,829 per task.
Granite 4.1 comprises three dense decoder-only models (3B, 8B, 30B) trained on ~15T tokens across five pre-training phases that progressively shift from broad web data to curated, domain-specific content.
Google describes TPUs as custom chips designed specifically to run AI models, emphasizing their ability to perform complex mathematical operations at scale.
Claude Code v2.1.122 introduces ANTHROPIC_BEDROCK_SERVICE_TIER environment variable for selecting default, flex, or priority Bedrock service tiers via X-Amzn-Bedrock-Service-Tier header.
OpenAI published an announcement detailing its safety framework for ChatGPT, covering model safeguards, misuse detection, and enforcement mechanisms.
AWS Bedrock now offers OpenAI's latest models, Codex, and a new Bedrock Managed Agents service designed for OpenAI reasoning models
Musk testified that SpaceX serves as 'life insurance' for humanity and that he co-founded OpenAI out of concern that AI could either save or destroy civilization
NVIDIA released Nemotron 3 Nano Omni, an open-weights multimodal model combining text, image, audio, and video understanding for document analysis, speech recognition, and agentic computer use.
Elon Musk began testimony in a California federal court trial he filed against OpenAI CEO Sam Altman and president Greg Brockman, along with Microsoft and OpenAI as co-defendants.
OpenAI announced its GPT models, Codex, and Managed Agents are now available on AWS infrastructure.
Transformers v5.7.0 added Poolside's Laguna mixture-of-experts model and DEIMv2 real-time object detection model to the library.
OpenAI and Microsoft announced a renegotiated agreement on April 27, 2026, replacing Microsoft's indefinite exclusive rights to all OpenAI products and IP with a nonexclusive license extending through 2032.
NVIDIA released NV-Raw2Insights-US, an AI model that learns directly from raw ultrasound sensor data rather than reconstructed images, enabling patient-specific sound-speed estimation for adaptive image focusing.
OpenAI announced it has achieved FedRAMP Moderate authorization status, a compliance credential required for federal agencies to use cloud services. The authorization covers both ChatGPT Enterprise and the OpenAI API. FedRAMP Moderate represents a specific security and risk management baseline established by federal standards.
OpenAI published a principles statement from CEO Sam Altman describing five core guidelines for the company's work on AGI development.
OpenAI has published Symphony, described as an open-source specification designed to orchestrate Codex-powered agents
OpenAI and Microsoft announced an amended partnership agreement
OpenAI released Privacy Filter, a 1.5B-parameter Apache 2.0-licensed model that detects eight categories of personally-identifiable information (PII) in a single 128k-token context window and achieves state-of-the-art performance on the PII-Masking-300k benchmark.
World Press Photo announced its 2026 winner, 'Separated by ICE' by Carol Guzy, while simultaneously releasing detailed rules about what qualifies as photography in competitions.
Meta is laying off approximately 10 percent of its workforce—roughly 8,000 employees—with notifications scheduled for May 20th, 2026.
Meta is deploying software called the Model Capability Initiative (MCI) on U.S. employee computers to capture mouse movements, clicks, keystrokes, and occasional screen snapshots for training AI agents.
Cohere is acquiring Germany-based Aleph Alpha with support from Schwarz Group, a German retail conglomerate, to create a European alternative to U.S.-dominated AI providers.
Google is investing up to $40 billion in Anthropic: $10 billion immediately at a $350 billion valuation, plus $30 billion contingent on hitting performance targets.
ComfyUI raised $30 million in new funding at a $500 million post-money valuation, led by Craft Ventures with participation from Pace Capital, Chemistry, and TruArrow.
DeepSeek released V4 in two MoE variants: V4-Pro (1.6T total, 49B active parameters) and V4-Flash (284B total, 13B active), both supporting 1M-token context windows
Claude Code v2.1.118 released April 23 with vim visual mode (v) and visual-line mode (V) for enhanced text selection and operators
TechCrunch confirmed that Delve performed security certifications for Context AI, which was linked to a breach affecting Vercel after an employee installed a malicious app and granted it access to Google-hosted corporate systems.
Shopify CTO Mikhail Parakhin participated in a Latent Space podcast interview covering internal AI adoption, tooling (Tangle, Tangent, SimGym), and how AI-assisted coding has shifted bottlenecks from code generation to review, testing, and deployment.
OpenAI announced GPT-5.5 on April 23, 2026, positioning it as a model built for complex, real-world work including coding, research, analysis, and document creation.
Sierra, the customer service AI startup founded by Bret Taylor and Clay Bavor, has acquired Fragment, a YC-backed French startup that helps businesses integrate AI into workflows.
Researchers presented COSPLAY, a framework where an LLM decision agent retrieves skills from a learnable skill bank while a parallel agent extracts reusable skills from unlabeled rollouts.
Researchers O'Herlihy and Català introduce the Defensibility Index (DI) and Ambiguity Index (AI) to evaluate content moderation systems based on policy consistency rather than agreement with human labels.
Google DeepMind introduced Decoupled DiLoCo, a distributed training architecture that divides compute into asynchronous 'islands' to improve resilience and reduce bandwidth in global data center training
Researchers at Harbin Institute of Technology document a widespread tendency for language models to call external tools unnecessarily, even when they possess adequate internal knowledge to answer questions.
ZeroFolio uses pretrained text embeddings instead of hand-crafted features to select algorithms across diverse problem domains including SAT, MaxSAT, QBF, ASP, CSP, MIP, and graph problems.
IBM chief HR officer Nickle LaMoreaux announced at Charter's Leading With AI Summit that the company plans to triple entry-level hiring this year, a move she characterized as 'provocative' given prevailing assumptions about AI workforce displacement
Meta has deployed a tool called the Model Capability Initiative (MCI) on US-based employees' work computers that captures mouse movements, clicks, keystrokes, and occasional screenshots to train its AI agents.
Qwen released Qwen3.6-27B, claiming it surpasses the Qwen3.5-397B-A17B across major coding benchmarks despite being 15x smaller (27B vs 397B parameters)
Researchers at Wrocław University of Environmental and Life Sciences published a study in Satellite Navigation describing how super-resolution generative adversarial networks (SRGAN) can transform blurry atmospheric readings into sharp humidity maps.
Gemma 4 VLA can run entirely locally on entry-level edge hardware (NVIDIA Jetson Orin Nano Super with 8GB RAM) with speech-to-text, text-to-speech, and webcam integration
Apple will present research at ICLR 2026 on parallel RNN training (ParaRNN), improved state space models with tool-use access, unified image understanding and generation (Manzano), 3D scene generation from photos, and protein folding approaches.
OpenAI announced a free ChatGPT program targeting verified U.S. healthcare professionals including physicians, nurse practitioners, and pharmacists
OpenAI published technical guidance on using WebSockets in its Responses API to speed up agentic workflows
Criminals are using generative AI to produce phishing emails, deepfakes, and malicious code at scale, with scam centers in Southeast Asia and state-level actors leveraging these tools.
OpenAI released GPT-5.2 in December and GPT-5.3 Codex in February, both trained on NVIDIA infrastructure including Hopper and GB200 NVL72 systems.
Allbirds, which sold its sneaker business for $39 million in March, rebranded as NewBird AI and announced $50 million in unnamed investor financing for GPU capacity, causing its stock to jump over 300%.
The S&P 500 has declined 3% this year, but underlying volatility is extreme: 57 stocks gained at least 20% while 47 fell at least 20%, according to Bespoke Investment Group analysis from mid-March 2026.
Apple researchers published a study examining information leakage from vision-language model internals, comparing what data survives compression through different representational bottlenecks
NVIDIA announced Rubin platform with 50 petaflops NVFP4 inference capacity, Vera CPUs, NVLink 6 networking, and AI-native storage claimed to deliver tokens at one-tenth prior-generation costs.
Apple researchers introduced MixAtlas, a framework for optimizing data mixtures in multimodal LLM pretraining, accepted at the NADPFM workshop at ICLR 2026.
NVIDIA's Graduate Fellowship Program, in its 25th year, is accepting applications worldwide for doctoral students working in AI, machine learning, autonomous vehicles, computer graphics, robotics, and related fields.
Anthropic updated claude.com/pricing on April 22, 2026, to move Claude Code access from the $20/month Pro plan to $100+ Max plans, but reversed this within hours.
John Ternus is succeeding Tim Cook as Apple CEO, with Cook remaining as executive chairman. The company faces multiple active antitrust disputes, including a DOJ case filed in March 2024 over smartphone market dominance, Epic Games litigation over App Store practices, and a potential $38 billion fine in India.
Meta is building an internal tool to capture employee keystrokes, mouse movements, and clicks for AI model training.
SpaceX announced a partnership with Cursor, the popular AI coding platform, that includes a buyout option valued at $60 billion, exercisable later in 2026.
GROVE is an interactive visualization system that displays multiple language model generations as overlapping paths through a text graph, exposing shared structure, branching points, and output clusters.
Apple researchers have published a peer-reviewed paper introducing a benchmark to systematically evaluate how well large language models understand context across four distinct tasks and nine datasets.
OpenAI announced the Codex Transformation Partners program, a joint initiative with major consulting and technology firms including Accenture, PwC, and Infosys to facilitate enterprise deployment of the Codex code-generation model.
Academic paper proposes Agentic AI Governance Maturity Model (AAGMM) with five levels spanning 12 governance domains grounded in NIST and ISO standards.
YouTube is extending its likeness detection tool—originally tested with creators in fall 2025 and expanded to politicians and journalists in March—to celebrities who can now flag and request removal of AI deepfakes.
YouTube has extended its AI-powered likeness detection tool beyond a limited creator pilot to serve the entertainment industry, including talent agencies and their represented celebrities, politicians, government officials, and journalists.
Filippo Valsorda published analysis arguing AES-128 provides adequate security even in a post-quantum computing environment, challenging the misconception that Grover's algorithm halves symmetric key strength
Salesforce has launched Headless 360, exposing its Salesforce, Agentforce, and Slack platforms as APIs and CLI tools for AI agents to access directly without a browser interface
Amazon has invested an additional $5 billion in Anthropic, bringing its total stake to $13 billion.
Honor's autonomous humanoid robot completed a 13.1-mile half-marathon in 50 minutes, 26 seconds on April 19, 2026, beating the human record of 57:20 set by Ugandan runner Jacob Kiplimo
NVIDIA published Nemotron-Personas-Korea, a 7-million-record synthetic persona dataset covering all 17 Korean provinces with 2,000+ occupations and ~209,000 unique names grounded in data from official sources including KOSIS and the Supreme Court of Korea.
Claude Code v2.1.116 achieves up to 67% faster resumption on sessions 40MB or larger, with optimized handling of dead-fork entries
Allbirds rebranded as an AI company and temporarily increased its stock price by roughly sixfold, exemplifying what observers describe as performative AI adoption.