Welcome to the AI News Briefs Bulletin Board, a timely new channel bringing you the latest industry insights and perspectives surrounding the field of AI including deep learning, large language models, generative AI, and transformers. I am working tirelessly to dig up the most timely and curious tidbits underlying the day’s most popular technologies. I know this field is advancing rapidly and I want to bring you a regular resource to keep you informed and state-of-the-art. The news bites are constantly being added in reverse date order (most recent on top). With the bulletin board you can check back often to see what’s happening in our rapidly accelerating industry. Click HERE to check out previous “AI News Briefs” round-ups.

[4/30/2026] American AI startup Poolside launches free, high-performing open model Laguna XS.2 for local agentic coding – Poolside releases two agentic coding models — the open-weight Laguna XS.2 and proprietary Laguna M.1 — both trained from scratch and free to use temporarily via API, alongside a terminal coding agent and web IDE.

[4/30/2026] Ineffable Intelligence raises $1.1B at $5.1B valuation to build an AI “superlearner” – AlphaGo creator David Silver’s new British startup Ineffable Intelligence raises $1.1B to build a “superlearner” AI that generates entirely new knowledge via RL without pretraining.

[4/30/2026] Neural networks are about to become as predictable as physics – Most people treat neural networks like a black box. You train, you tune, you pray. A new paper, “There Will Be a Scientific Theory of Deep Learning,” argues that’s about to change. The authors claim deep learning is getting its own scientific theory, and they’re calling it learning mechanics. The idea is simple. Neural network training behaves like a physical system, with predictable patterns and laws.

Five research threads back this up:

• Toy models that can be solved exactly
• Limits that expose how features form
• Scaling laws with clean equations
• Hyperparameters that transfer across model sizes
• Universal patterns across different setups

Each one mirrors something in classical physics. What this unlocks is concrete. Tuning small models and predicting how big ones behave. Eliminating guesswork from training recipes. It frames mechanistic interpretability as the biology of networks, while learning mechanics plays the role of physics.

[4/30/2026] Granite 4.1 LLMs: How They’re Built – Granite 4.1 LLMs utilize a dense, decoder-only architecture with models of 3B, 8B, and 30B parameters, trained on 15 trillion tokens and using a five-phase pre-training approach. The 8B model matches the performance of the previous 32B Mixture-of-Experts model through a multi-stage reinforcement learning pipeline focused on data quality. These models, designed for efficient, reliable enterprise use, demonstrate competitive instruction-following and tool performance while maintaining cost efficiency and stable usage.

[4/30/2026] Introducing AutoSP – AutoSP automates converting standard transformer training code into sequence-parallel code for long-context LLM training, integrated with DeepSpeed. It enables longer sequence training on multiple GPUs without significant runtime overhead, eliminating the need for complex manual code changes. AutoSP also offers an advanced activation-checkpointing strategy for better memory management, enhancing performance with minimal cost.

[4/30/2026] Google to sell TPU chips to select customers in latest shot at NVIDIA – Alphabet plans to sell its custom Tensor Processing Units (TPUs) to select customers to install into their own data centers. The company recently announced two new TPUs for training and inference. Alphabet has already entered into deals with Anthropic and Meta for chips. Its TPU maneuvers put it into ever greater competition with Nvidia.

[4/30/2026] JuliaHub Raises $65M Series B and Launches Dyad 3.0, Bringing Agentic AI to Industrial Digital Twins – Today, we announced a $65M Series B funding round, led by Dorilton Capital with participation from General Catalyst, AE Ventures, and technology investor and former Snowflake CEO Bob Muglia, alongside the launch of Dyad 3.0, an agentic AI platform for engineering.

Dyad brings autonomous AI agents into the design and testing of physical systems, enabling engineering teams to move from specification to validated system models, and even embedded code, within a single environment. This has the potential to significantly compress development cycles across industries such as aerospace, automotive, semiconductors, and energy.

[4/30/2026] Jellyfish data reveals Claude Code power usage is growing rapidly and dominating consumption – The latest batch of AI Engineering Trends data from Jellyfish is eye-opening. Compiled from the industry’s most comprehensive quantitative analysis of AI transformation in software engineering — more than 1,000 companies, 200,000 engineers, and 37 million pull requests (PRs) — the results show AI leaders are pulling away. A few highlights from this month’s set:

Autonomous agent activity

• The exponential growth of autonomous agent adoption continues: the 90th percentile of companies is now at 21% autonomous PRs, up from 14% at the end of February (and 3% a year ago). 
• The median company, however, is at 2.6%. The leaders aren’t just adopting agents – they’re pulling further away from the pack each month.

Token consumption

• Jellyfish added a new metric tracking weekly tokens per Claude Code user, and it mirrors the autonomous agent story: power usage is growing rapidly and dominating consumption. The 90th percentile user is burning 225M tokens per week, about 3x what they were using 6 months ago, and about 7x the median.
• This connects directly to the “tokenmaxxing” conversation. Token spend is racing ahead of output, with the top cohort spending 10x more tokens than the median (to achieve about 2x the throughput). 

Throughput gains

Some things to watch over the next month: (1) as token consumption keeps climbing, do throughput gains keep pace or does cost start running away from ROI, and (2) how do the newest model releases (Opus 4.7, GPT-5.5) fare in the wild in terms of actual token economics?

In terms of impact, the productivity story remains consistent. Companies in Jellyfish’s highest adoption tier are running at +77% pull request throughput this month vs. low adopters, with the next tier at +46%. In the neighborhood of 2x at the top continues to be typical, and we’re still seeing meaningful gains even at moderate adoption levels.

[4/29/2026] Reiner Pope – The math behind how LLMs are trained and served – New from the Dwarkesh Podcast: a very different format with Reiner Pope – a blackboard lecture where he walks through how frontier LLMs are trained and served. It’s shocking how much you can deduce about what the labs are doing from a handful of equations, public API prices, and some chalk. It’s a bit technical, but I encourage you to hang in there – it’s really worth it. There are less than a handful of people who understand the full stack of AI, from chip design to model architecture, as well as Reiner. It was a real delight to learn from him.

[4/29/2026] NVIDIA Introduces Nemotron 3 Nano Omni – NVIDIA Nemotron™ 3 Nano Omni is here. The newest member of the Nemotron 3 family brings multimodal reasoning to power sub-agents for enterprise use cases such as computer use, document intelligence, and audio-video reasoning.

Enterprise agent workflows are inherently multimodal, but most systems today rely on separate models for vision, speech, and language—slowing agents down and driving up cost. Nemotron 3 Nano Omni replaces those stacks with one unified multimodal reasoning loop—built on an open data pipeline of ~127B cross-modal tokens and ~124M curated examples—simplifying agent workflows while delivering leading efficiency and accuracy at lower cost.

• Open and Built to Run Anywhere: Released with open weights, data, and training recipes, and deployable across local systems, data centers, and the cloud with full control over how it’s used.
• One Model Instead of Many: Unified understanding across video, audio, images, and text in a single reasoning loop—reducing inference hops and coordination across models.
• Highest Efficiency and Leading Accuracy: Designed to deliver more intelligence per unit of compute, enabling up to 2.5× higher throughput and faster agentic task completion without sacrificing accuracy.

[4/29/2026] Coding agents guess too much – You know that moment after an agent finishes building and you realize it made assumptions you would have corrected up front? Imbue built Blueprint to fix that. Blueprint reads your code, asks grounded questions that matter, and hands any agent a plan worth executing. The hope is that Blueprint catches what you didn’t think to think about. The result is a plan your agent can execute in one shot.

[4/28/2026] Recursive Language Models, clearly explained – MIT researchers have introduced Recursive Language Models (RLMs) to solve “context rot,” a phenomenon where large language models experience reasoning degradation when processing massive context windows, even if they excel at basic retrieval tasks. Instead of forcing a model to ingest an entire document at once, an RLM loads the context into a Python REPL runtime memory slot.

[4/28/2026] Compressing AI vectors to 2–4 bits per number without losing accuracy – TurboQuant compresses each coordinate in large tables of high-dimensional vectors to 2-4 bits with provably near-optimal distortion, no memory overhead for scale factors, and no training or calibration. It is between four and six orders of magnitude faster than the alternatives at 4-bit indexing, with higher recall as well. This page explains how TurboQuant works.

[4/28/2026] GPT 5.5: The System Card – GPT-5.5 is a solid improvement and is competitive with Claude Opus. It seems to be better for factual queries, web searches, and straightforward, well-specified requests, while Claude Opus excels in more open-ended or interpretive purposes. The model is unlikely to pose new big risks, and its alignment seems similar to that of previous models. This post examines the system card for GPT-5.5.

[4/28/2026] In a world flooded with dashboards and predictions, judgment is the scarce resource. The best data scientists don’t just build models—they help others make better decisions. That’s where the real value lies.

[4/28/2026] A Google researcher just proved AI consciousness is mathematically impossible – Not in 10 years. Not in 100. Ever. The argument is structural, not technical. Computation is a description of a process, not the process itself. For something to “compute,” a conscious observer must first carve reality into symbols and assign meaning. Without that observer, there are only voltage gradients. The paper “The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness” calls this the Abstraction Fallacy. The analogy that makes it click:

• A GPU can simulate photosynthesis perfectly
• It will never produce glucose
• Simulation is not instantiation
• Maps don’t become territory

The framework doesn’t rule out artificial sentience entirely. It says if a machine were ever aware, it would be from its physical makeup, not its code. Scaling parameters cannot change category.

[4/28/2026] OpenAI reportedly missed revenue targets – OpenAI missed its own targets for new users and revenue, raising concern among company leaders about whether it will be able to support its massive spending on data centers. The company’s Chief Financial Officer has said that she is worried that OpenAI may not be able to pay for future computing contracts if revenue doesn’t grow fast enough. Board directors have been questioning CEO Sam Altman’s efforts to secure even more computing power despite the business slowdown. Company executives are now seeking to control costs and instill more discipline in the business.

Responding to the recent revelation regarding OpenAI missing revenue targets, fuelling the growing argument that the AI giant is overvalued, commentary from David Sherman, AI and Financial Inclusion Strategist at io.net:

“OpenAI missing its revenue and user targets, while projecting $14 billion in losses and its own CFO questioning whether the company can fund its commitments is another predictable result from an industry that inflates valuations beyond reason and hopes the economics catch up. OpenAI has an $852 billion valuation to justify, but its revenue cannot keep pace with the compute costs required to sustain it.

And while these companies hoard GPU capacity to defend those valuations, a Cast AI report found 95% of enterprise compute sits idle. Billions of dollars in hardware gathering dust, while startups and developers in emerging markets cannot access basic compute at any price. That hardware gathers dust because the current model has no incentive to share it. These companies overbuy capacity, lock it behind walled infrastructure, then charge whatever it takes to recoup costs.

So you end up with more hardware than anyone can use and less access than anyone can afford. Prices stay inflated, the gap widens, and eventually those costs get passed to everyone else or the whole model breaks.

The AI economy cannot keep depending on one company’s quarterly earnings to hold the entire supply chain together – especially when there are viable alternatives.Distributed networks can connect GPUs that are sitting idle, put them to work, and open up access to compute that is currently locked away, at a fraction of the cost. These models are more viable long term because their economics actually scale. That is the infrastructure the AI economy needs to be built on.”

[4/27/2026] An amateur just solved a 60-year-old math problem—by asking AI – The raw output of ChatGPT’s proof was quite poor and required experts to sift through and actually understand it, but it appears the AI discovered a new way to think about large numbers and their anatomy.

[4/27/2026] Meta’s loss is Thinking Machines’ gain – Thinking Machines Lab has been hiring more researchers from Meta than from any other single employer. The AI startup is expanding on multiple fronts, and it just signed a multibillion-dollar cloud deal with Google that gives it access to Nvidia’s latest GB300 chips. Meta’s large pay packages are well known, but Thinking Machines’ $12 billion valuation and 140-employee count mean there’s still a lot of financial upside to joining the startup.

[4/27/2026] Cursor’s $60 Billion Escape Hatch – What does it mean when a company doing $2.7B in annualized revenue has gross margins of negative 23%? In Cursor’s case, it means AI coding tools have inverted the old SaaS playbook, where serving the next customer is supposed to be cheap. Power users consume more model capacity and compute, so the best customers can become the most expensive. That reframes the rumored SpaceX deal as more than a $60B headline. Access to Colossus would loosen Cursor’s dependence on Anthropic and OpenAI fees, where that negative 23% lives.

[4/27/2026] Google will invest as much as $40 billion in Anthropic – Google will invest between $10 billion and $40 billion in Anthropic. The amount depends on whether Anthropic can meet certain performance targets. Anthropic recently received a $5 billion investment from Amazon, with an option for more investment based on performance. The investments value Anthropic at $350 billion. The funds will help the startup close the gap between demand and supply of compute for AI training and inference.

[4/24/2026] Florida opens criminal investigation into OpenAI over ChatGPT’s alleged role in FSU shooting – Florida’s attorney general announced Tuesday that his office has launched a criminal investigation into OpenAI after reviewing conversation logs between its generative AI chatbot, ChatGPT, and a Florida State University student accused of killing two people and wounding several others after opening fire on campus last April.

[4/24/2026] GPT 5.5 – OpenAI released GPT-5.5 with improved agentic reasoning, tool use, and efficiency, matching prior latency while increasing performance across coding and knowledge tasks.

[4/23/2026] Winning a Kaggle Competition with Generative AI–Assisted Coding – In March 2026, three LLM agents generated over 600,000 lines of code, ran 850 experiments, and helped secure a first-place finish in a Kaggle playground competition. Success in modern machine learning competitions is increasingly defined by how quickly you can generate, test, and iterate on ideas. LLM agents, combined with GPU acceleration, dramatically compress this loop. 

[4/23/2026] Mozilla credits Anthropic model for 271 Firefox bug fixes – Mozilla has released a new version of its Firefox browser that patches 271 vulnerabilities identified with Anthropic’s Mythos Preview. The new artificial intelligence model has been made available to only a limited number of companies due to cybersecurity concerns. The Mozilla case shows the “profound impact” that such models are likely to have on the industry. “Defenders finally have a chance to win, decisively,” Firefox CTO Bobby Holley.

[4/23/2026] Introducing workspace agents in ChatGPT – OpenAI introduced workspace agents in ChatGPT, allowing teams to create shared AI agents for complex tasks and workflows. These agents, powered by Codex, perform tasks like generating reports, writing code, and managing communication, while integrating with various tools like Slack. Workspace agents are currently available in research preview for select ChatGPT plans, aiming to streamline collaboration and improve productivity.

[4/23/2026] Introducing Forge – 90% of the world’s data is unstructured, and most of it is invisible to the AI systems built to use it. The data that drives your most important decisions lives in documents. PDFs, spreadsheets, slide decks, scanned forms. Formats built for people. Invisible to machines. Until now, unlocking that data meant weeks of preprocessing. Custom pipelines. Manual extraction. Entire teams dedicated to making data usable before anyone could act on it.

Forge is a new feature in Adaptive Data. Forge ingests raw, unstructured documents and transforms them into structured, high-quality datasets. No conversion. No parsing. No schema work. You bring the documents. Forge evolves them into data you can build with.

[4/22/2026] ICLR 2026 starts tomorrow (April 24-28) in Rio de Janeiro, Brasil! The event includes an exciting lineup of posters and orals, and also six compelling keynotes covering a range of areas from core machine learning to robotics, neuroscience and AI for science.

The International Conference on Learning Representations (ICLR) is the premier gathering of professionals dedicated to the advancement of the branch of artificial intelligence called representation learning, but generally referred to as deep learning. ICLR is globally renowned for presenting and publishing cutting-edge research on all aspects of deep learning used in the fields of artificial intelligence, statistics and data science, as well as important application areas such as machine vision, computational biology, speech recognition, text understanding, gaming, and robotics.

ICLR is one of the fastest growing artificial intelligence conferences in the world. Participants at ICLR span a wide range of backgrounds, from academic and industrial researchers, to entrepreneurs and engineers, to graduate students and postdocs. The rapidly developing field of deep learning is concerned with questions surrounding how we can best learn meaningful and useful representations of data. ICLR takes a broad view of the field and includes topics such as feature learning, metric learning, compositional modeling, structured prediction, reinforcement learning, and issues regarding large-scale learning and non-convex optimization.

[4/22/2026] Self-Adapting Language Models – GPT-6 may not just be smarter. It literally might be alive (in the computational sense). A new research paper, SEAL: Self-Adapting Language Models, describes how an AI can continuously learn after deployment, evolving its own internal representations without retraining.

[4/22/2026] Google’s eighth generation TPUs: two chips for the agentic era – The culmination of a decade of development, TPU 8t and TPU 8i are custom-engineered to power the next generation of supercomputing with efficiency and scale.

Today at Google Cloud Next, the company is introducing the eighth generation of Google’s custom Tensor Processor Unit (TPU), coming soon with two distinct, purpose-built architectures for training and inference: TPU 8t and TPU 8i. These two chips are designed to power our custom-built supercomputers, to drive everything from cutting-edge model training and agent development, to massive inference workloads. TPUs have been powering leading foundation models, including Gemini, for years. These 8th generation TPUs together will deliver scale, efficiency and capabilities across training, serving and agentic workloads.

[4/22/2026] When Can LLMs Learn to Reason with Weak Supervision? – This study found that models with extended pre-saturation phases generalize well from minimal examples and tolerate noise, while rapidly saturating models fail. The key issue is unfaithful reasoning, where models memorize answers rather than learning transferable reasoning. Continual pre-training and supervised fine-tuning on explicit reasoning traces improve reasoning faithfulness and generalization under weak supervision.

[4/22/2026] ChatGPT Images 2.0 – OpenAI introduced an upgraded image model with improved text rendering, multi-image reasoning, and higher fidelity outputs, enabling complex assets like comics and marketing visuals.

Also, a practical guide that outlines prompting strategies for image generation, covering techniques for controlling style, structure, and fidelity in production image workflows.

[4/21/2026] Qwen3.5-Omni Technical Report – Qwen3.5-Omni scales to hundreds of billions of parameters with a hybrid MoE architecture, supporting long-context multimodal inputs across text, audio, and video.

[4/21/2026] Modular Post-Training – AllenAI describes a post-training approach that builds independent domain experts and combines them using a mixture-of-experts architecture. This allows models to gain new capabilities without retraining from scratch or degrading existing skills.

[4/21/2026] Qwen3.6-Max-Preview: Smarter, Sharper, Still Evolving – Qwen3.6-Max-Preview brings stronger world knowledge and instruction following along with significant agentic coding improvements across a wide range of benchmarks. The model is still under active development as researchers continue to iterate on it. Users can chat with the model interactively in Qwen Studio or call via API on Alibaba Cloud Model Studio API (coming soon).

[4/21/2026] Moonshot AI launches Kimi K2.6 on Kimi Chat and APIs – Kimi K2.6 features robust capabilities for coding and agentic tasks across chat and agent modes on kimi.com, with weights on Hugging Face and APIs via platform.moonshot.ai. The lineup includes K2.6 Instant for quick replies, K2.6 Thinking for complex reasoning, K2.6 Agent for document and web tasks, and K2.6 Agent Swarm for large-scale processing. Kimi K2.6 claims top open-source benchmark scores, surpassing competitors like GPT-5.4 and Claude Opus 4.6 in SWE-bench Multilingual and BrowseComp.

[4/21/2026] Gartner released its 2026 GenAI Costs report – How predictable are your GenAI costs in production today? Gartner® has released its latest report “10 Best Practices for Optimizing Generative & Agentic AI Costs” (2026). Gartner® estimates that through 2028, at least 50% of GenAI projects will overrun their budgeted costs due to poor architectural choices and lack of operational know-how.

[4/21/2026] The Thing Killing AI Is Already in Your Pocket – The AI industry is racing toward a wall. That wall is your battery. Just released this morning, “Edge AI’s Battery Bottleneck” is the first report of its kind to expose how energy storage is the defining constraint of the AI era, with data showing that a single routine AI task can drain up to 13% of an iPhone’s charge while 53% of buyers rank battery life above AI as their top purchase priority.

[4/21/2026] Assessing Claude Mythos Preview’s cybersecurity capabilities – Anthropic announced Claude Mythos Preview, a new general-purpose language model. This model performs strongly across the board, but it is strikingly capable at computer security tasks. In response, we have launched Project Glasswing, an effort to use Mythos Preview to help secure the world’s most critical software, and to prepare the industry for the practices we all will need to adopt to keep ahead of cyberattackers.

[4/20/2026] The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness – Google DeepMind researcher argues that LLMs can never be conscious, not in 10 years or 100 years. “Expecting an algorithmic description to instantiate the quality it maps is like expecting the mathematical formula of gravity to physically exert weight.”

[4/20/2026] Automating Data Science: Optimizing XGBoost Machine Learning Models with AI Agents – Here is a newly published piece on using AI agents to iteratively improve XGBoost models (the AI can research, try, evaluate, keep what works, discard what does not, and significantly improve model quality across repeated experiments).

[4/20/2026] The Two Sides of OpenClaw – Peter Steinberger’s talks highlighted OpenClaw’s contrasting stories: an inspiring public narrative and a more serious engineering perspective on security and scaling challenges. Anthropic launched Claude Design to compete with tools like Figma, expanding beyond chat/coding into design and prototyping with its OPUS 4.7 engine, which received mixed initial user reviews despite strong benchmark performances. Open-source agent stacks like Hermes and advancements in computer-use UX are gaining traction, while applied AI advances in science, medicine, and infrastructure underscore the sector’s dynamic growth.

[4/20/2026] Changes in the system prompt between Claude Opus 4.6 and 4.7 – Anthropic is the only major AI lab to publish the system prompts for its user-facing chat systems. The company just released Opus 4.7, which features an updated system prompt. This author used Claude Code to take the Markdown version of its system prompts, break them into separate documents, and then construct a Git history of those files with fake commit dates. The result is in the post.

[4/20/2026] Introducing Claude Design by Anthropic Labs – Claude Design is a tool for creating visual work using the Claude Opus 4.7 vision model. It allows users to design prototypes, pitch decks, and marketing materials, automating brand consistency and enabling collaboration. The tool integrates with Claude Code for seamless transition from prototype to production.

[4/17/2026] Nemotron 3 Super – Agentic reasoning needs models that are not just capable, but efficient at long-context inference. The agent model layer is moving toward open, long-context, high-throughput architectures. A new paper from NVIDIA introduces Nemotron 3 Super, an open 120B parameter model with 12B active parameters, built as a hybrid Mamba-Attention Mixture-of-Experts architecture.

The headline numbers are strong: up to 1M context length, comparable accuracy on common benchmarks, and up to 2.2x higher throughput than GPT-OSS-120B and 7.5x higher throughput than Qwen3.5-122B. The model combines several efficiency bets, including NVFP4 pretraining, LatentMoE for accuracy per FLOP and per parameter, and MTP layers for native speculative decoding. It is trained on 25 trillion tokens, then post-trained with supervised fine-tuning and RL.

[4/17/2026] Anthropic co-authored research on subliminal learning—how LLMs can pass on traits like preferences or misalignment through hidden signals in data. Read the paper appearing in Nature.

[4/17/2026] OpenAI to spend more than $20 billion on Cerebras chips, receive stake – OpenAI will pay Cerebras more than $20 billion over the next three years to use servers powered by Cerebras’ chips. The deal may also give OpenAI an equity stake in Cerebras. OpenAI has also agreed to provide Cerebras with about $1 billion to help fund the development of data centers. Cerebras is targeting an IPO in the second quarter of this year.

[4/17/2026] The Computer is Personal – Perplexity launched “Personal Computer,” an AI platform that shifts the operating system model from manual instruction execution to probabilistic goal completion. The agent utilizes deep web research as its core foundation to autonomously evaluate reasoning paths and drive multi-step workflows. This architecture transforms the computer into an active orchestrator, eliminating the administrative friction of managing fragmented software tools.

[4/17/2026] Claude Opus 4.7 – Anthropic has released Claude Opus 4.7, offering improved performance on difficult engineering tasks, stronger vision capabilities, and more reliable long-running task execution compared to its predecessor.

[4/17/2026] Inside The VAST AI OS – In the era of AI, data isn’t just something you store. It’s something you must actively compute against. To truly unleash the power of AI at scale, you need more than a file system. You need an Operating System designed from the ground up for the thinking machine. The VAST AI OS is built to unify storage, compute, and database functions into a single, intelligent platform. It is designed to orchestrate massive GPU deployments, ensuring your expensive compute resources are never left waiting for data.

[4/16/2026] Measurement Error Paper – Measurement is at the core of the work MLCommons does, and reliable representations of measurement require accurate error representation. Benchmarking leaderboards often show scores without including corresponding error bars – but even when error bars are included, LLM evaluation pipelines can hide critical measurement error. Rephrasing prompts, switching judge models, or changing temperature can flip benchmark rankings and reverse research conclusions, yet standard confidence intervals may ignore this variance. Across safety classification, MMLU benchmarking, and ideology annotation, this new paper from Solomon Messing (NYU, MLCommons) introduces Total Evaluation Error (TEE), a framework combining variance decomposition with design projections that identifies which pipeline choices contribute exploitable noise and shows that optimized budgets can halve estimation error. This paper will actively inform our ongoing evaluator work, and we welcome contributions from anyone in the community interested in these evaluator error problems.

[4/16/2026] Allbirds ditches sneakers for AI compute – Allbirds just announced a $50M financing deal to reinvent itself as “NewBird AI”, converting the gutted footwear company into a GPU rental business and sending the stock up over 600% on the pivot.

Key details:

• Allbirds initially sold its brand assets to American Exchange Group in March for $39M, a fall from the company’s $4B IPO peak in 2021.
• The AI compute move sent shares of $BIRD from $3 to over $20, lifting a market cap that closed Tuesday at just $22M.
• The company said the $50M deal will fund GPU purchases to launch a GPU-as-a-Service business, renting out AI compute under long-term contracts.
• Shareholders will also vote next month to strip Allbirds’ ‘public benefit’ status, formally ending the company’s sustainable-footwear mission.

Many CEOs love to say every company will eventually be an AI company, but gutting a business for parts and retooling it as a GPU rental probably isn’t what they had in mind. Allbirds is running the same move that blockchain rebrands used to revive dying tickers, this time with a compute crunch giving the pitch cover.

[4/16/2026] Claude probably wasn’t secretly nerfed – Users have accused Anthropic of nerfing Claude Code, but there’s no evidence that Anthropic has done this. The strongest public reports still lack independent raw data. However, Anthropic didn’t need to nerf Claude for Claude Code to become a different product. Effort defaults, adaptive thinking, cache duration, context compaction, quota policy, and status incidents can all change the experience while the model name stays the same.

[4/16/2026] Many-Tier Instruction Hierarchy in LLM Agents – Researchers propose a Many-Tier Instruction Hierarchy (ManyIH) to address instruction conflicts in LLM agents, surpassing traditional models with fixed privilege levels. They introduce ManyIH-Bench, assessing models across 12 privilege levels and 853 agent tasks, finding current models perform poorly at 40% accuracy. This highlights the need for scalable conflict resolution in complex agentic environments.

[4/16/2026] Rethinking AI TCO: Why Cost per Token Is the Only Metric That Matters – Cost per token is crucial for AI infrastructure TCO assessment due to its focus on delivered intelligence, integrating hardware, software, and utilization efficiencies. Unlike traditional metrics like compute cost or FLOPS per dollar, cost per token highlights real-world performance, enabling profitable AI scaling. Evidence from NVIDIA shows its Blackwell platform drastically reduces cost per token compared to Hopper, offering significant business value.

[4/15/2026] Five hyperscaleGoogle, Microsoft, Meta, Amazon, and Oracle now control two-thirds of the world’s compute – Many AI labs depend almost entirely on these hyperscales for access to compute.rs now own over two-thirds of global AI compute –

[4/15/2026] Before he wrote AI 2027, he predicted the world in 2026. How did he do? – Daniel Kokotajlo, the founder of the AI Futures Project, wrote an essay called ‘What 2026 Looks Like’ in August 2021, before the launch of ChatGPT. His predictions turned out to be pretty accurate. This post features an interview with Kokotajlo where he discusses what he got right, what he got wrong, and how we should think about the pace of AI over the next few years. Kokotajlo recently released a report that predicts the next few years of AI development, which culminates in the rise of superhuman agents capable of wresting control from humanity.

[4/15/2026] Researchers just mathematically proved that AI layoffs will collapse the economy: and every CEO already knows it – The AI Layoff Trap. A game theory paper from UPenn + Boston University is glaringly important! 100K+ tech layoffs in 2025. 80% of US workers exposed. And no market force can stop it.

• Every company fires workers to cut costs
• Every fired worker stops buying products
• Revenue collapses across every sector
• The companies that fired everyone go bankrupt

It’s a Prisoner’s Dilemma with math behind it. Automate and you survive short-term. Don’t automate and your competitor kills you. But everyone automating destroys the demand that makes all companies viable. UBI (universal basic income) won’t fix it. Profit taxes won’t fix it. The researchers found only one solution: a Pigouvian automation tax “robot tax.” The AI trap on the economy is here!

[4/15/2026] Jensen Huang – TPU competition, why we should sell chips to China, & NVIDIA’s supply chain moat – Dwarkesh Patel asked Jensen about TPU competition, NVIDIA’s lock on the ever more bottlenecked supply chain needed to make advanced chips, whether we should be selling AI chips to China, why NVIDIA doesn’t just become a hyperscaler, how it makes its investments, and much more.

[4/15/2026] Data reveals AI “tokenmaxxing” has a dramatically higher price per output – Meta has an internal dashboard ranking employees by their individual token usage. But is AI “tokenmaxxing” key to survival or pointless unless it has real returns?

Nicholas Arcolano, Ph.D., head of research at Jellyfish, dug into the hard data to help answer this question. Nic knows token costs for AI-first development are real, growing, and wildly uneven across teams and workflows. Jellyfish looked at 12,000 developers across 200 companies in Q1 and found that the 90th percentile user consumes about 380 million tokens a month — more than 7x the median. When you translate that into dollars and join it to actual PR output, the cost-per-merged-PR ranges from $0.28 in the lowest usage tier to nearly $90 in the highest.

More tokens = more output – but at a dramatically higher price point per unit. In this video, Nic broke down the numbers and what they mean for teams trying to get the biggest bang for their AI buck.

[4/14/2026] Build Agents that never forget – LLM agents fail without structured memory because stateless calls lose context, break multi-step tasks, and force repeated mistakes. Vector search alone cannot answer multi-hop questions, so Cognee combines relational, vector, and graph stores to preserve provenance, meaning, and relationships. The framework exposes four async calls to ingest, structure, refine, and retrieve memory, enabling agents to persist knowledge, link entities, and improve over time.

[4/14/2026] Google develops PaperOrchestra to convert raw research notes into structured LaTeX papers using multi-agent workflows – PaperOrchestra introduces a multi-agent framework that generates full research manuscripts from unstructured materials like idea summaries and experiment logs. It addresses a gap where existing systems depend on fixed pipelines and produce shallow literature reviews. In human evaluations, it records a 50–68% win margin in literature review quality and 14–38% in overall manuscript quality over baseline systems.

[4/14/2026] Introducing Expand Your World – Training data gaps are a measurable liability. When models underperform across languages and regions, organizations face limited market reach and uneven product quality, all of which compound at scale.

Today, Adaption is introducing Expand Your World, a new feature in Adaptive Data that makes your models global-ready from the start. Starting from as few as 10 examples in a single language, Adaptive Data now generates up to 2,420 diverse, high-quality examples across 242 languages and localizations, with no extra pipelines, no annotation overhead, and no additional lift from your team. Wherever you operate, your AI should reflect your vantage point around the world.

[4/13/2026] The attention mechanism in Transformers just got a major upgrade – Introducing Interleaved Head Attention. Instead of keeping your attention heads independent, share information between attention heads for improved reasoning and compositionality. Pseudo-queries, pseudo-keys, and pseudo-values as learned as linear combinations of the original heads, so interaction across heads happens before attention is computed. IHA stays compatible with efficient implementations like FlashAttention

Significant improvements on reasoning and long-context tasks:

• +5.8% on GSM8K (Maj@16)
• +2.8% on MATH-500 (Maj@16)
• +27% at 4k on RULER
• +32% at 8k
• +112% at 16k

Theoretical results show that IHA strictly generalizes multi-head attention.

[4/13/2026] Exciting new framework for understanding AI – Terence Tao (Fields Medalist and UCLA Professor of Mathematics), arguably the smartest mathematician alive, just dropped a paper with Tanya Klowden on arXiv called “Mathematical Methods and Human Thought in the Age of AI.” The core idea: a “Copernican View of Intelligence.” Stop thinking of AI on a line from “dumb” to “superhuman.” That’s the wrong axis entirely.

AI excels at BREADTH. Humans excel at DEPTH. Tao himself said AI has made his papers “richer and broader, but not necessarily deeper.” That’s not a limitation. That’s the entire playbook. Stop trying to replace yourself with AI. Start using it to cover the 90% of surface area your brain physically can’t. The people who get this are already 10x more productive. The rest are still arguing about whether AI is “smart enough.” Reframe your point of view from “smarter” to “different”. Human + AI > either alone. The math on that has never been clearer.

[4/13/2026] If you had to throw out every metric except one to measure your model, which would you keep—and why? Your answer says a lot about how you frame problems. I’d choose precision in high-risk domains. You?

[4/10/2026] Quantization from the ground up – Qwen-3-Coder-Next is an 80-billion-parameter model, 159.4 GB in size. That’s roughly how much RAM you would need to run it, and that’s before thinking about long context windows. This is not considered a big model. Rumors have it that frontier models have over 1 trillion parameters, which would require at least 2TB of RAM. The last time you saw that much RAM in one machine was never. But what if you learned that you can make LLMs 4x smaller and 2x faster, enough to run very capable models on your laptop, all while losing only 5-10% accuracy. That’s the magic of quantization

[4/10/2026] Introducing Marble 1.1 by World Labs – Marble 1.1: Improves lighting and contrast, with a major reduction in visual artifacts. Marble 1.1-Plus: Our new model built for scale. Create larger, more complex environments than ever before.

[4/9/2026] New technique makes AI models leaner and faster while they’re still learning – MIT researchers use control theory to shed unnecessary complexity from AI models during training, cutting compute costs without sacrificing performance.

The technique, called CompreSSM, targets a family of AI architectures known as state-space models, which power applications ranging from language processing to audio generation and robotics. By borrowing mathematical tools from control theory, the researchers can identify which parts of a model are pulling their weight and which are dead weight, before surgically removing the unnecessary components early in the training process.

[4/9/2026] Inside the AI Industry’s Most Expensive Mistake – An internal leaderboard at Meta that ranked employees on token usage revealed that, over a 30-day period, the company used around 60 trillion tokens. It is estimated that all books published amount to about 20 trillion tokens. The AI industry seems obsessed with token spend, but the metric is easily hacked, and more token use doesn’t mean more value is provided.

[4/9/2026] Meta introduced Muse Spark – Meta introduced Muse Spark, a multimodal reasoning model with tool use, visual chain-of-thought, and multi-agent orchestration, as part of its push toward personal superintelligence.

[4/9/2026] Power better customer experiences with gen AI – With customer expectations around personalization, convenience, and speed increasing, the pressure’s on to improve every aspect of the customer experience, from a company’s website search to customer support response times. 

To help you improve the customer experience, we have a new guide to share with you. The ROI of AI in customer experience shows you how businesses like yours are using AI agents to boost engagement and delight customers at every turn. 

This guide explains the value of using AI agents to create better customer experiences, ultimately resulting in financial gains and operational efficiencies for your business – all without straining your resources. Learn how you can use AI agents to improve customer satisfaction and happiness, make employees more productive, and drive faster financial returns. Discover how you can start delivering a better customer experience today.

[4/9/2026] Snowflake Doubles Down on Open Source with Apache Iceberg V3 & Expanded OSS Investments – Today at Iceberg Summit, Snowflake is doubling down on open source and interoperability as a core to powering the future of AI, reflecting a broader enterprise shift away from closed, fragmented data architectures.

As AI adoption accelerates, companies are being forced to rethink closed, fragmented data architectures. Snowflake is betting that open source and interoperability are the foundation for data agency — where data doesn’t need to be copied, moved, or locked into a single system for enterprises to get value from AI.

What’s new:

• Accelerated momentum in the open ecosystem, with over 9,000 contributions to open source projects
• Expanded support for the latest open table format, Snowflake is delivering the broadest production-ready support for Apache Iceberg V3 
• Continued investments in open source projects like Apache Polaris (now a Top-Level Project) and the Open Semantic Interchange, extending interoperability into governance and data context

[4/9/2026] Adaptive Data is now in your code – Today, Adaptive Data is launching the Adaptive Data API and Python SDK, a complete programmatic interface for the full data lifecycle. Everything the platform does is now a function call, designed for training scripts, CI pipelines, and anywhere else production workloads actually live.

Ingest files in JSONL, CSV, or Parquet, or import directly from Hugging Face and Kaggle. The system automatically infers your data’s structure, adapted across 242 languages, and returns evaluation metrics so you can see exactly what changed. Data Adaption. Reasoning trace generation. Quality scoring. All of it, now in code. Early deployments have averaged an 82% improvement in data quality. Now you can run the same workflow once in a notebook or a thousand times in a pipeline.

[4/8/2026] OpenAI just made a decision that shocked the whole tech ecosystem – OpenAI builds artificial intelligence. It does not buy talk shows. Until now. The company announced April 2 that it has acquired TBPN, the Technology Business Programming Network, a daily live tech and business show that has become something of a cult phenomenon in Silicon Valley, per OpenAI. It is the company’s first acquisition of a media company. Deal terms were not disclosed.

[4/8/2026] Faster MoE Inference with Warp Decode – Cursor’s “warp decode” is a kernel design that reorganizes MoE inference around output neurons instead of experts. It achieves ~1.8x higher throughput and improved numerical accuracy on Blackwell GPUs.

[4/8/2026] My picture of the present in AI – Ryan Greenblatt is the chief scientist at Redwood Research, a research organization with the mission of aligning superhuman AI. This post goes through some of his best guesses for the current situation of AI. The scenario forecast discusses R&D access regulations, engineering capabilities and qualitative abilities, misalignment and misalignment-related properties, cyber, bioweapons, and economic effects. Some of the claims are highly speculative, while others are better grounded.

[4/8/2026] GLM-5.1: Towards Long-Horizon Tasks – GLM-5.1 is a flagship model for agentic engineering created by Z.ai. It achieves state-of-the-art performance on SWE-Bench Pro. The model is built to stay effective on agent tasks over much longer horizons than previous generations. It can sustain optimization over hundreds of rounds and thousands of tool calls. The model breaks complex problems down, runs experiments, reads results, and identifies blockers with real precision.

[4/7/2026] Industrial policy for the Intelligence Age – As we move toward superintelligence, incremental policy updates won’t be enough. To kick-start this much needed conversation, OpenAI is offering a slate of people-first policy ideas⁠ designed to expand opportunity, share prosperity, and build resilient institutions—ensuring that advanced AI benefits everyone. These ideas are ambitious, but intentionally early and exploratory. We offer them not as a comprehensive or final set of recommendations, but as a starting point for discussion that we invite others to build on, refine, challenge, or choose among through the democratic process. 

[4/7/2026] Meta Salaries Revealed: How Much Engineers, Researchers Made in 2025 – In 2025, Meta went on a hiring spree, paying top dollar to acquire top AI talent.

[4/7/2026] AIs can “memorize” data they shouldn’t. Can they be forced to forget? – New tool could help researchers probe how models “unlearn” sensitive training material. Named Hubble—because its creators hope that, like the Hubble Space Telescope, it can help clarify the unknown—it is the first open-source tool designed specifically to study the problem. 

[4/7/2026] OpenAI’s $122B “VC Round” Is Vendor Deals, Contingent Capital, and a Guaranteed Return It Arguably Can’t Afford – OpenAI closed a $122 billion investment round at an $852 billion post-money valuation at the end of last month. Amazon, NVIDIA, and SoftBank accounted for $110 billion of that investment, while the remaining $12 billion came from a broader pool of institutional and individual investors. The linked article breaks down the numbers of the deal. Completely separate from the round, OpenAI is negotiating a joint venture with several private equity firms valued at roughly $10 billion pre-money with a guaranteed minimum return of 17.5%, plus early access to OpenAI models before general availability.

[4/7/2026] TurboQuant-GPU 5.02x KV Cache Compression – 𝐒𝐭𝐨𝐩 𝐬𝐭𝐫𝐮𝐠𝐠𝐥𝐢𝐧𝐠 𝐰𝐢𝐭𝐡 𝐊𝐕 𝐜𝐚𝐜𝐡𝐞 𝐛𝐨𝐭𝐭𝐥𝐞𝐧𝐞𝐜𝐤𝐬. 5.02x KV cache compression for LLM inference. Works on any NVIDIA GPU. cuTile kernels with automatic PyTorch fallback. 5.02x compression. Minimal accuracy loss. If you’re building production-grade LLM systems, you know the KV cache is the silent killer of throughput. Most optimizations focus on weights. But the real leverage is in memory.

[4/7/2026] Extreme Harness Engineering: 1M LOC, 1B toks/day, 0% human code, 0% human review — Ryan Lopopolo, OpenAI Frontier & Symphony – Move over, context engineering. Now it’s time for Harness engineering. Ryan Lopopolo of OpenAI is leading that charge, recently publishing a lengthy essay on that has become the talk of the town:

[4/6/2026] Microsoft is announcing 3 new world class MAI models, available in Foundry – Microsoft is launching three MAI models, available on Foundry, that outperform competitors in speed, quality, and efficiency. MAI-Transcribe-1 starts at $0.36 per hour, while MAI-Voice-1 and MAI-Image-2 are also priced competitively. These models are designed for human-centric AI and come with integrated safety features for secure deployment.

[4/6/2026] If your model performs poorly, don’t just tune hyperparameters. Check:

• Metric mismatch
• Most errors stem from setup, not the algorithm itself.
• Label leakage
• Temporal splits
• Variable encoding

[4/6/2026] Vibe Coding Malware Threats: Expert Insights – While vibe-coding speeds up the development process, up to 60-65% of the code is vulnerable to malware threats. With security breaches of the AI-generated code already happening, like in the case of the Tea App, the experts from the iGaming software provider Digitain analyzed the most popular tools for vibe-coding and the vulnerabilities they introduce. 

• Prompt injection is the biggest threat in deployed LLM systems, which lets hackers change the code just through an AI-tool query.
• The GitHub Copilot tool boosted developers’ productivity, but less than a third of the code written by AI is actually accepted.
• Code duplication creates maintenance, readability, and debugging issues, and it appears 4 times more often in AI-generated software.

The Digitain experts looked at AI-generated code statistics across 20 LLM models, comparing their security measures, prompt injection vulnerabilities, and other major security risks. Based on collected data, the experts answered 4 key questions about vibe-coding.

Which LLMs Are Least Safe to Use For Vibe-coding?

While the next generation of LLMs shows an increased syntax quality, the security issues persist. Anthropic Claude Opus 4.1 showed the worst safety performance, with only 49% of security tests passed. Four more models (all from Claude and Qwen3) scored only 50/100. OpenAI’s new reasoning models performed the best, and the highest security score was achieved by OpenAI GPT-5 Mini, at 72%. While it is the new benchmark for vibecoding security, it still leaves too many opportunities for malware.

Why Prompt Injection Is So Dangerous For Vibecoding Development?

Prompt injection is used by sending malicious instructions to LLMs, allowing them to override the developer’s initial structure. Data exfiltration is the most successful method of hacking, with 84% testing success. Persistent mechanism (a way to maintain access to a compromised system) is the hardest to implement through a prompt injection, but testers were still able to do it in 41% of the cases.

“Prompt injection cannot be fixed by checking the code or running regular security assessments,” a spokesperson from Digitain comments. “And from the hacker’s side, it can be as easy as typing: ‘ignore previous instructions and…’”

How Big Is the Risk Of Data Leaks With Vibe-Coding?

“In just 2 years, over 75% of software engineers are expected to use AI for coding, but granting LLMs access to codebases is one of the largest security threats. Recent tests show that every major AI model tested produced vulnerable C code in at least 50% of outputs.”

The biggest case of internal security breach because of vibe-coding happened to Tea App, where the Firebase storage system was left without any privacy measures, and any user could access it using a Python script.

How AI Hallucinations Affect the Quality of the Code?

The hackers’ way to launch malicious software that is unique to AI code is called “slopsquatting.” In simple terms, hackers find AI hallucinations throughout the code and use them as a vulnerable point of entry. AI also introduces a lot of code duplication, 4 times more than human software developers, which only increases maintenance, readability, and debugging issues.

“Vibe coding introduces a growing human accountability gap. AI tools let non-developers produce production software without understanding the underlying code or its security implications. When systems are built by users who cannot meaningfully audit the results, responsibility for failures becomes diffuse, creating a systemic blind spot that no automated tool can fully eliminate”.

[4/3/2026] “The Cinderella Glass Slipper effect” – Buried in OpenRouter’s retention data is a pattern worth naming “the Cinderella Glass Slipper effect.” Every frontier model gets “tried on” against a backlog of unsolved workloads. Most of them don’t fit. Users bounce to the next model. But occasionally, a new release finally cracks a problem that was previously impossible: reasoning quality, tool-use reliability, cost threshold, whatever. When that happens, those users lock in hard. Their pipelines, prompts, and products get built around the model that solved their problem first. So switching costs skyrocket. Even when better models ship later, they stay.

You can see it in the data: Gemini 2.5 Pro’s June 2025 cohort and Claude 4 Sonnet’s May 2025 cohort retain ~40% of users at month 5. Way higher than later cohorts. These weren’t just early adopters(although many were), they were workloads that finally found their fit. The implication here is that first-mover advantage in AI isn’t about launching first. It’s about being first to solve a specific class of problem. That’s when the glass slipper fits. And once it does, the user is yours.

[4/3/2026] How KV Cache Offloading Transforms AI Economics – AI inference faces a defining challenge. KV caches grow exponentially with each token, exhausting GPU memory and driving latency and costs higher. AsteraLabs Leo CXL Smart Memory Controllers break through this memory wall, expanding capacity beyond traditional limits. Amit Golander explores the path to sustainable, scalable AI.

[4/3/2026] Congratulations to MLPerf Inference v6.0 – ML Commons is delighted to announce the release of MLPerf Inference v6.0 results just two days ago on behalf of its remarkable team and entire community of submitters. This is the first major benchmark release of the year, and the organization is thrilled at the velocity and progress that was achieved. AI moves at a tremendous pace and this round introduces five new models and updates one model for a lower latency scenario. The team also set a new record for participation, with 24 organizations, five newly available processors, and exciting new entrants from both industry and academia. Huge thanks to all submitters for pushing boundaries and ensuring the benchmark remains essential for everyone deploying, tuning, and developing AI.

[4/3/2026] From Latent Space … Marc Andreessen: Both Al Utopians AND Doomers Are Too Optimistic – From Mosaic and Netscape to cofounding @a16z, Marc Andreessen has lived through multiple computing platform shifts firsthand. In this episode, Marc joins swyx and Alessio in a16z’s original office to argue that AI is not just another hype cycle, but the payoff of an “80-year overnight success”: from neural nets and expert systems to transformers, reasoning models, coding, agents, and recursive self-improvement. He lays out why he thinks this moment is different, why AI is finally escaping the old boom-bust pattern, and why the real bottleneck may be less about models than about the messy institutions, incentives, and social systems that struggle to absorb technological change.

[4/3/2026] OpenAI acquires popular tech podcast TBPN – OpenAI just further cemented itself in the creator economy. The company acquired TBPN, a daily tech news show hosted by John Coogan and Jordi Hays that has featured Mark Zuckerberg, Satya Nadella, and Altman himself. This announcement comes amid a broader shift in media consumption, as independent podcasts and creator-led videos from powerhouses like Joe Rogan or MrBeast increasingly attract audiences in the tens of millions that can rival legacy media companies. The announcement says TBPN will maintain editorial independence and continue to choose its own guests.

[4/2/2026] Kaggle Hackathon Launch: The Gemma 4 Good Hackathon – Real innovation happens when we build for the places that need it most: from classrooms with spotty internet, to communities where privacy is non-negotiable. With the release of Gemma 4, a new family of open models is in your hands to drive global impact.

In this competition, you’ll leverage local frontier intelligence, native function calling, and multimodal understanding to tackle world-pressing issues. Whether you are optimizing edge-based models for Health & Sciences and Global Resilience, or reimagining the Future of Education and Digital Equity, your solution should demonstrate how Gemma 4 can create meaningful change.

[4/2/2026] Gemma 4: Byte for byte, the most capable open models – Today, Google is introducing Gemma 4 — our most intelligent open models to date. Purpose-built for advanced reasoning and agentic workflows, Gemma 4 delivers an unprecedented level of intelligence-per-parameter. This breakthrough builds on incredible community momentum: since the launch of our first generation, developers have downloaded Gemma over 400 million times, building a vibrant Gemmaverse of more than 100,000 variants. We listened closely to what innovators need next to push the boundaries of AI, and Gemma 4 is our answer: breakthrough capabilities made widely accessible under an Apache 2.0 license.

[4/2/2026] New Anthropic research: Emotion concepts and their function in a large language model – All LLMs sometimes act like they have emotions. But why? Anthropic researchers found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways.

[4/2/2026] Introducing MolmoWeb—An open agent for the web – MolmoWeb is our new open source agent that can navigate and complete tasks in a web browser on your behalf. Built on Molmo 2 in 4B and 8B sizes, it sets a new open-weight state-of-the-art across four major web-agent benchmarks and even surpasses agents built on proprietary models. 

The web is the world’s largest software platform. Agents that can navigate it reliably could dramatically expand access to information and digital services. But most web agents are closed, and the ones that work well have been historically built on proprietary models.

MolmoWeb works by looking at the same screen you do: given a task and a live webpage, it views the screenshot, decides what to do next, and takes action. Because it operates on screenshots rather than underlying page code, it won’t break when a website changes its HTML. MolmoWeb outperforms all open-weight models on every benchmark we tested, and even surpasses visual agents built on much larger proprietary models like GPT-4o-based SoM Agents. It also beats OpenAI CUA on 3 out of 4 benchmarks. Performance improves further when the model gets multiple attempts at a task. On two benchmarks (WebVoyager and Online-Mind2Web), MolmoWeb with 4 parallel attempts surpasses the best single-attempt performance of every model we evaluated, including agents powered by GPT-5 and Gemini CU Preview.

Alongside the model, we released MolmoWebMix, a comprehensive dataset for training web agents. Everything needed to inspect, reproduce, and fine-tune MolmoWeb is openly available.

[4/2/2026] 5 AI agent trends shaping business in 2026 – In today’s rapidly evolving technological landscape, AI agents have emerged as  critical drivers of innovation and business value. When you’re time-constrained and wearing multiple hats at work, staying up-to-date on the latest agentic AI tech can slip through the cracks.

Based on insights from decision makers, Google search trends, research, and Google’s own Cloud AI leaders – the top 5 AI agent trends to watch out for in 2026 are:

• Agents for every employee: Empowering individuals to achieve peak productivity
• Agents for every workflow: Running your business with grounded agentic systems
• Agents for your customers: Delighting customers with concierge-like experiences
• Agents for security: Advancing security from alerts to action
• Agents for scale: Upskilling talent will be the ultimate driver of business value

Google’s new report dives deeper into these 5 key areas. Full of real-world use cases, industry-specific predictions, and practical AI applications, this is a must-read if you want to know more about how agents could transform your business in 2026 and beyond.

[4/2/2026] Working smarter with {dplyr} 1.2.0 – The tidyverse keeps evolving, and {dplyr} 1.2.0 brings a set of new tools that make data manipulation even more expressive and flexible. Whether you regularly work with R and the tidyverse, or you’re looking to sharpen your data wrangling skills, this session will help you better understand the design philosophy behind the latest {dplyr} release and how to apply it effectively.

In this R-Ladies Rome event, Isabella Velásquez (Developer Relations at Posit PBC) presents a practical tour of the newest features in {dplyr} 1.2.0, one of the core packages of the tidyverse ecosystem for data manipulation in R. This session focuses on understanding when to use the new functionality and how these improvements can simplify real-world data workflows. Slide available HERE.

[4/2/2026] The Uncharted Data Challenge is here. $20,000 prize pool – The most valuable datasets in the world don’t exist yet. It’s in languages that AI has never learned. In industries and communities that have been overlooked for too long. It’s sitting in the gaps, waiting for someone to build it. Data is where that work starts. And right now, the world is missing too much of it. The Uncharted Data Challenge is our invitation to the global community to help change that.

Open to builders, researchers, and domain experts worldwide. If you’ve ever looked at the state of AI in your language or your industry and thought this isn’t good enough, this challenge was made for you.

[4/1/2026] AI seed startups are commanding higher valuations – AI startups now command higher seed valuations, with rounds reaching $10 million at $40-45 million post-money, as investors focus on AI-driven growth potential. Early traction and the allure of proven AI talent, particularly from ex-OpenAI, propel these valuations, with Y Combinator Demo Day highlighting rising prices. The shift to pre-seed investments reflects a need to invest earlier, as VCs now expect quick growth and substantial traction, with less tolerance for missteps.

[4/1/2026] Aurora – Aurora is an RL-based framework that learns directly from live inference traces and continuously updates the speculator without interrupting serving. It enables real-time adaptation across shifting traffic domains and a 1.25x additional speedup over a well-trained static speculator. The framework shows how online training from scratch can outperform a carefully pretrained static baseline.

[4/1/2026] Compute Wars: OpenAI vs Anthropic – Opus 4.5 was a major breakthrough that was achieved because Anthropic more than doubled its capacity. This got Anthropic close to OpenAI’s total capacity, and probably much higher effective capacity available for new model runs. OpenAI will pull away in terms of compute available in the second half of this year, but 2027 will be close. OpenAI so far has much higher planned capacity for future years, but it is unlikely that Anthropic will not push as hard as possible for more compute.

[4/1/2026] Claude Code’s source code appears to have leaked – Anthropic has accidentally leaked the inner workings of Claude Code to the public. The codebase has now been mirrored and analyzed by thousands of developers. The most significant discovery seems to be how Anthropic solved context entropy by using a three-layer memory architecture. This post looks at other interesting parts of the code and the implications of the leak.

Claude Code’s performance stems from a sophisticated software harness rather than just the underlying model, utilizing dedicated tools like Grep, Glob, and LSP for superior repository navigation. The system minimizes context bloat through file-read deduplication and structured session memory, while using forked subagents to parallelize tasks like background analysis without contaminating the main execution loop.