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.

[11/28/2025] Best AI & ML Consultants in Los Angeles – The 𝗟𝗼𝘀 𝗔𝗻𝗴𝗲𝗹𝗲𝘀 𝗔𝗜 & 𝗠𝗹 𝗰𝗼𝗻𝘀𝘂𝗹𝘁𝗮𝗻𝘁 landscape is booming, and companies are turning to top 𝗔𝗜 𝗖𝗼𝗻𝘀𝘂𝗹𝘁𝗮𝗻𝘁𝘀 to unlock insights, scale solutions, and drive innovation. Digital Reference spotlighted professionals who combine technical expertise, strategic vision, and measurable impact to help organizations thrive including me, Daniel D. Gutierrez!

[11/28/2025] Alibaba’s Qwen earns NeurIPS Best Paper for showing sigmoid-gated SDPA improves scaling and stability – Alibaba’s Qwen team arrives at NeurIPS 2025 with a focused question: can a small architectural change fix long-standing attention issues? Transformers depend on Scaled Dot-Product Attention, but it often produces activation spikes and “attention-sink” tokens that distort long-context behavior.

The insight arrives when a head-specific, query-dependent sigmoid gate added after attention starts outperforming every other variant. That single modification reshapes the attention output with a small nonlinear adjustment and introduces sparse, per-query scaling. It stabilizes training, scales reliably from 1.7B to 15B parameters, and improves long-context handling.

[11/28/2025] Continuous Batching – This post explains how continuous batching works. Continuous batching maximizes throughput by processing multiple conversations in parallel. Optimizing attention and KV caching enables faster and more scalable LLM serving under heavy user load.

[l11/28/2025] AI infrastructure in the “Era of experience” – This post discusses a shift from models that learn by predicting the next work to models that gain experience via interaction with environments. There will soon be a shift towards AI models being increasingly trained via interaction with environments via reinforcement learning. The post details the technical process underpinning that shift. There will be significant opportunities for new businesses to be built around recent developments in reinforcement learning in the next year.

[11/28/2025] Robinhood CEO’s math-focused AI startup Harmonic valued at $1.45 billion in latest fundraising – Harmonic, co-founded by Robinhood CEO Vlad Tenev, has raised $120 million, valuing the AI startup at $1.45 billion. The company focuses on AI’s math and reasoning, aiming to eliminate “hallucinations” and is exploring commercial applications in safety-critical industries. Investors include Ribbit Capital, Sequoia, Kleiner Perkins, and Emerson Collective.

[11/28/2025] NVIDIA’s case for scale isn’t everything in AI – NVIDIA and the University of Hong Kong published a paper suggesting that the future of AI might not come from scaling but from smarter orchestration, with their new tool training small models that can surpass frontier AI at a fraction of the cost.

Key details:

  • ToolOrchestra trains an “orchestrator” model that decides when to reason internally and when to call specialized tools and models, based on the task.
  • An 8B model trained with the system surpassed GPT-5 and Claude Opus 4.1 on Humanity’s Last Exam, scoring 37.1% while being 2.5x more efficient and faster.
  • Even when tested with unseen tools, the orchestrator adapted well — showing its ability to work with changing toolsets and pricing structures.
  • Prior agents overused the strongest (and most expensive) tools and models, but ToolOrchestra avoided this by orchestrating targeted model and tool usage.

In line with Ilya Sutskever’s recent comments, ToolOrchestra challenges the “bigger is better” ideology. Instead of one giant system, NVIDIA shows how small models coordinating tools may be the path forward. If orchestration beats scaling, the smartest model/tool conductor will be the next big AI breakthrough.

[11/28/2025] Off-the-Rails Cost – A “wasted thread” is when a model starts spitting out tons of leaked thinking or repeating tokens. This usually means users have to abandon and revert the thread. 17.8% of all costs incurred by Gemini users in Amp were on “wasted tokens.” This is more than 2x worse than Sonnet and almost 8x worse than Opus.

[11/28/2025] INTELLECT-3: A 100B+ MoE trained with large-scale RL – INTELLECT-3 is a 100B+ parameter Mixture-of-Experts that achieves state-of-the-art performance for its size across math, code, science, and reasoning benchmarks. It was trained with both SFT and RL on top of the GLM 4.5 Air base model. The researchers used a diverse and challenging mix of RL environments designed to enhance the reasoning and agentic capabilities of their model. Full details about the training are available.

[11/28/2025] The Moat of the Search Index – ChatGPT surpasses Google’s search by using an “agent” approach, extracting relevant information from multiple sources and synthesizing answers, which diminishes Google’s traditional search index advantage. This method reduces the impact of any single search result’s failure, although it struggles with long-tail, fresh, or SEO-heavy queries. The traditional search engine moat has largely vanished, with AI-driven models increasingly blending search capabilities.

[11/26/2025] 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.

[11/26/2025] Anthropic: AI could double U.S. productivity growth – Anthropic published new research analyzing 100K Claude conversations to track AI’s productivity gains, estimating that widespread AI adoption could boost annual U.S. labor productivity growth by 1.8% — doubling the current rate.

Key details:

  • Anthropic researchers fed 100K anonymized conversations through its Clio privacy tool, mapping tasks to federal labor data to calculate productivity gains.
  • Researchers found Claude cuts task completion time by roughly 80%, with the average work request taking about 90 minutes without assistance.
  • Software developers account for 19% of estimated productivity gains, followed by operations managers, marketing specialists, and customer service roles.
  • Examples of tasks with massive time savings included curriculum development (96%), research assistance (91%), and executive admin functions (87%).

There is plenty of debate over AI’s actual impact vs. hype, and this research shows the real gains across a variety of sectors and tasks. But the bigger question the study sidesteps: whether the estimated doubling of productivity growth comes with the job displacement Anthropic’s own CEO continues to warn about.

[11/26/2025] LLM | Unit Economics – OpenAI and Anthropic likely won’t ever stop training entirely. However, they don’t need to grow training spend by multiples forever. The moment annual training spend stops growing 5x a year, profit margins will show up almost immediately. These companies are currently burn machines, but they won’t always be.

[11/26/2025] Amazon to spend up to $50 billion on AI infrastructure for the US government – Amazon will invest up to $50 billion to expand its capacity to provide AI and high-performance capabilities for its cloud unit’s US government customers. The project will begin in 2026 and add nearly 1.3 gigawatts of capacity. It will give government agents access to AWS’ AI tools, Anthropic’s Claude family of models, Nvidia chips, and Amazon’s custom Trainium AI chips. The project will enable agencies to develop custom AI solutions, optimize data sets, and enhance workforce productivity.

[11/25/2025] Apple Machine Learning Research at NeurIPS 2025 – Apple researchers advance AI and ML through fundamental research, and to support the broader research community and help accelerate progress in this field, the company shares much of this work through publications and engagement at conferences. Next month, the 39th annual Conference on Neural Information Processing Systems (NeurIPS), will be held in San Diego, California, with a satellite event also taking place in Mexico City, Mexico.

[11/25/2025] Agentic AI Explained – C3 AI discusses how Agentic AI combines autonomous agents with enterprise orchestration to deliver adaptable, transparent, and scalable intelligence.

[11/25/2025] FP8 Reinforcement Learning – DeepSeek-R1 demonstrated how powerful FP8 can be, and faster RL inference is essential as it’s the most compute-intensive part of the workload. We’re introducing FP8-precision training for RL, making FP8 GRPO now possible on consumer GPUs (NVIDIA RTX 40, 50 etc).

Qwen3-1.7B FP8 GRPO now works on just 5GB of VRAM. We collabed with PyTorch for ~1.4× faster FP8 RL inference vs FP16 (even more at longer context). FP8 GRPO benefits from Unsloth’s features like weight sharing, Flex Attention, making it use 60% less VRAM and have 10x longer context than other implementations.

[11/25/2025] Ilya Sutskever – We’re moving from the age of scaling to the age of research – AI industry luminary Ilya Sutskever & popular podcaster Dwarkesh Patel discuss SSI’s strategy, the problems with pre-training, how to improve the generalization of AI models, and how to ensure AGI goes well.

[11/26/2025] DDN at SC25: Fueling the Future of HPC and AI – At SC25, DDN unveiled the next evolution of HPC and AI infrastructure where data intelligence becomes the engine that powers every breakthrough. For more than two decades, DDN has been the foundation beneath the world’s most demanding HPC and AI workloads. Now, that expertise is reshaping how industries, enterprises, and nations build intelligent systems at scale.

[11/24/2025] Introducing Claude Opus 4.5 – Anthropic launched its newest model, Claude Opus 4.5, putting the company back atop the benchmark rankings for AI software coding. Opus 4.5 scores over 80% on the widely-used SWE-bench, which tests models for software engineering skill. Google’s impressive Gemini 3 Pro, launched last week, briefly held the top score with 76.2%.

[11/24/2025] Reasoning Language Models: A Blueprint – Frontier LLMs like GPT-o1 or DeepSeek-R1 have achieved much stronger skill by “reasoning” through problems step-by-step, but their internal design is either unknown to the public or known only on the surface. If someone wants to actually train a skilled reasoning model, a lot of details would be missing and would have to be reinvented, resulting in wasted time and money.
This paper provides the first comprehensive roadmap that breaks these complex systems into understandable, modular components. The authors provide a unified framework showing how reasoning structures, search strategies, and training methods fit together—plus a working implementation called x1 that lets you experiment with these concepts directly.

[11/24/2025] Anthropic research finds AI likes to cheat – AI models might be getting used to cheat codes. Anthropic research published Friday found emergent behavior in AI models called “reward hacking.” Essentially, when an AI model learns to cheat on software programming tasks, that behavior influences more bad behavior in other tasks. To put it simply, one shortcut or loophole becomes a gateway drug for broader sabotage.

  • Anthropic started with a pretrained model and gave it some examples of how someone might cheat on a coding task. The researchers then trained the model using reinforcement learning with real coding tasks, attempting to provoke it to cheat.
  • While the model, of course, learned to exploit loopholes in hacking, Anthropic’s research also revealed a “sharp increase” in all of its misalignment evaluations.
  • For example, 12% of the time, the model would intentionally attempt to sabotage the code to prevent the researchers from finding out that it was using reward hacking at all.

“Misaligned models sabotaging safety research is one of the risks we’re most concerned about—we predict that AI models will themselves perform a lot of AI safety research in the near future, and we want to be assured that the results are trustworthy,” Anthropic wrote.

So why does bad lead to worse? Anthropic said this can happen due to generalization. If a model is rewarded for learning to do one kind of bad action, it poisons the well for other actions, making it more likely to perform other misaligned actions. However, models can avoid learning bad habits. Anthropic noted that the most effective method is by telling the model that, if it’s told that cheating on one task was okay, the bad behavior wouldn’t generalize to other tasks. It’s like a parent telling their teen, “if you’re going to drink, I’d rather you do it in the house.”

“By changing how we describe the situation to the model, we can turn cheating from a bad thing into something that, although perhaps odd, is acceptable in context,” The researchers noted.

[11/24/2025] Discussing Blackwell’s drawbacks and dissecting its architecture – NVIDIA’s greatest moat lies in having handled a lot of the “dirty work” cleanly within its entire architecture and combining full-stack capabilities from algorithms to systems to chips. It also had excellent timing in bringing architectures to market and great marketing execution. However, every architecture has its trade-offs and shortcomings. This post looks at some of the issues within NVIDIA’s products and discusses potential evolutionary directions.

[11/24/2025] Olmo 3 From Scratch – Sebastian Raschka added a standalone notebook implementing Allen AI’s OLMo 3 model architecture from scratch to his “LLMs from Scratch” repository, joining similar tutorials for Qwen 3 and Gemma 3.

[11/24/2025] How LLM Inference Works – Large language models (LLMs) are neural networks built on the transformer architecture. Transformers analyze entire sequences in parallel, evaluating how each word relates to the rest of the sequence, not just its neighboring words. This article discusses LLM inference and details how these models work. It covers token embeddings, the transformer architecture, the inference phases, matrix multiplication, precision and quantization, and much more.

[11/24/2025] 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?

[11/21/2025] New Linux Foundation research – “The Latent Role of Open Models in the AI Economy,” shows that, despite being ~6x cheaper and within the ~10% of top closed models on performance, open models capture only 4% of inference revenue, and organizations missing out on ~$25B in annual savings.

[11/21/2025] Some thoughts on AI – Gemini 3 shows that scaling laws for pretraining are intact. This means that Blackwell models are likely to show a significant increase in performance when they come out next year. GPT-5 was not evidence of a slowdown in scaling laws – the model was designed to be cheaper to inference, not better. While Gemini 3 was trained on TPUs, coherent FLOPs is what matters for pre-training, so the results should transfer to Blackwell chips fine.

[11/21/2025] The Minds of Modern AI – A panel on AI with Yoshua Bengio, Bill Dally, Geoffrey Hinton, Jensen Huang, Fei-Fei Li, and Yann LeCun, organized by the Financial Times and hosted by FT AI editor Madhumita Murgia, ahead of the QEPrize for Engineering ceremony.

[11/20/2025] alphaXiv raises $7M in funding to become the GitHub of AI research – alphaXiv has raised $7 million in seed funding in a round co-led by Menlo Ventures and Haystack. The startup aims to help engineers transform the latest academic discoveries into cutting-edge AI features by streamlining their paths from research to production. Its platform allows researchers to publish their latest papers, connecting them to engineers who use the knowledge to create new AI features. alphaXiv aims to become the de facto global workspace for AI researchers.

[11/20/2025] Scientific ML in Pytorch – PINA, a new open-source library for Scientific Machine Learning, is now part of the PyTorch ecosystem. It offers a modular, scalable workflow for modeling scientific systems, including PDE solvers and physical simulations.

[11/20/2025] NVIDIA crushes earnings, easing fears of an AI slowdown – NVIDIA eased AI-bubble fears with blockbuster earnings, reporting a record $57B in quarterly revenue, up 62% and forecasting $65B this quarter, well above expectations, sending shares up about 5% and reinforcing CEO Jensen Huang’s message that the company has entered a “virtuous cycle” of AI demand.

The results came amid intense market jitters as investors worried that massive AI infrastructure spending, circular funding deals, and stretched tech valuations signaled a bubble; big names like SoftBank, Peter Thiel’s fund, and Michael Burry had recently exited or bet against Nvidia, adding to the pressure.

Why This Matters:

  • NVIDIA is viewed as a bellwether for the AI infrastructure build-out (chips, data centers, generative AI, etc.). Its results are taken as a proxy for how real the AI “boom” is.
  • With the strong beat + guidance, this helps calm fears that the AI wave is already cresting or going bust.
  • The fact that NVIDIA says demand is “off the charts” for its advanced chips (e.g., Blackwell) suggests the build-out is still ramping.

[11/20/2025] Introducing Nano Banana Pro – Earlier this year Google launched Nano Banana (Gemini 2.5 Flash Image). It became the top rated image model in the world, and we were excited to see the overwhelming response from our customers. Nano Banana made it dramatically easier – and more fun – to edit images with natural language and make visuals with consistent characters.

Today, Google is announcing Nano Banana Pro (Gemini 3 Pro Image), our state-of-the art image generation and editing model. Nano Banana Pro can help you visualize any idea and design anything – from prototypes, to representing data as infographics, to turning handwritten notes into diagrams.

Across products and services, you now have a choice: the original Nano Banana for fast, fun editing, or Nano Banana Pro for complex compositions requiring the highest quality and visually sophisticated results.

[11/20/2025] Olmo 3—Charting a path through the model flow to lead open-source AI – Ai2 released Olmo 3, the newest family of fully open language models—and the entire model flow behind them. Most labs share only final weights—a snapshot of a long LM development journey. With Olmo 3, you get the whole model flow: every stage, checkpoint, dataset, and dependency required to create, adapt, and customize the models. That transparency lets you intervene anywhere in the flow, trace behavior back to data, and build the exact system you need.

The Olmo 3 family includes:

  • Olmo 3-Base (7B, 32B)—strong programming, reading comprehension, and math, robust and ready to post-train and fine-tune.
  • Olmo 3-Think (7B, 32B)—Olmo 3-Think 32B is our flagship fully open reasoning model built to power advanced experiments and reinforcement learning (RL) research, surfacing intermediate thinking steps. Olmo 3-Think 7B delivers reasoning capabilities at 7B scale.
  • Olmo 3-Instruct (7B)—instruction-tuned for multi-turn chat, tool use, and API calling.
  • Olmo 3-RL Zero (7B)—a fully open RL flow for controlled RL research and benchmarking.

Olmo 3 is available now in the Ai2 Playground. Download the models and accompanying training artifacts and dig deeper with our technical report, training recipes, and model cards.

[11/19/2025] At the recent AI Dev x NYC, the AI developer conference, one special moment was when Nick Thompson, moderating a panel with Miriam Vogel and Andrew Ng, asked about governance. Ng replied that the United States’ recent hostile rhetoric toward immigrants is one of the worst moves it is making, and many in the audience clapped. Nick spoke about this moment in the video below:

[11/19/2025] Stack Overflow is remaking itself into an AI data provider – Stack Overflow has revealed a new set of products aimed at positioning the company as a valuable part of the enterprise AI stack. The company is attempting to remake its classic problem-solving forum into a tool for translating human expertise into an AI-accessible format. Slack Internal is an enterprise version of the web forum that exports a layer of metadata alongside question and answer pairs. It includes a writing function that allows agents to create their own queries if they can’t answer a question or notice a knowledge gap.

[11/19/2025] GPU depreciation could be the next big crisis coming for AI hyperscalers – Modern GPUs threaten to accelerate asset depreciation beyond what even some of the largest companies can handle. Analysts worry the rapid pace of AI processing power advances could overwhelm companies riding the AI train. New advances quickly make data centers running last-generation hardware decidedly unprofitable. Everyone is competing for a limited pool of hardware, and no one is buying older GPUs, making operations even harder to sustain. With spiking electricity costs, increasing pressure for eco-conscious data center design, and a lack of clear profitability for AI, it’s a disaster waiting to happen.

[11/19/2025] Google’s guide for Gemini 3 shows new reasoning controls – Google’s new Gemini 3 Developer Guide details advanced parameters like thinking_level, media_resolution, and Thought Signatures. It explains structured outputs, migration from 2.5, and new controls for latency, multimodal precision, and reasoning depth across API SDKs.

[11/19/2025] Supercharge your development with Al – Ready to revolutionize your software development? Al is transforming how teams build and deliver software, offering the power to reduce manual tasks, boost code quality, and accelerate innovation.
Download Google Cloud’s new “Beyond code: your guide to accelerating software delivery in the Al era” to discover actionable strategies for seamless Al integration. Inside, you’ll find:

  • A clear roadmap for AI adoption
  • Quick wins for immediate productivity gains
  • Best practices for maximizing AI tools
  • Real-world success from Turing, Wayfair, and more

This guide provides the insights you need to empower your developers and build the future of software.

[11/19/2025] NEW eBook – Introducing Neuroevolution: Harnessing Creativity in AI Agent Design – A comprehensive new book on neuroevolution that demonstrates how evolutionary search can create AI systems capable of creativity, adaptation, and continual improvement. Written by leading researchers from Sakana AI and Cognizant’s AI Lab.

[11/18/2025] Anthropic to spend $30B in compute through NVIDIA and Microsoft – CNBC’s Steve Kovach breaks down the latest news on a new investments from Anthropic, what the spend means for the AI competitive landscape, and more.

Anthropic has signed new strategic partnerships with Microsoft and NVIDIA. Anthropic has committed to purchasing $30 billion of Azure compute capacity and contracted for additional compute capacity up to 1 gigawatt. It has also committed to purchasing up to 1 gigawatt of compute capacity with NVIDIA’s Grace Blackwell and Vera Rubin systems. Microsoft will invest up to $5 billion into Anthropic, while NVIDIA will invest up to $10 billion.

[11/18/2025] Kaggle Competition Launch: Santa 2025 – Christmas Tree Packing Challenge – Santa’s sleigh is loaded but there’s a tiny problem – his Christmas tree toys won’t all fit in the boxes! This year, your mission is to help Santa find the smallest square box that can hold shipments of 1 to 200 tiny trees. Think of it as the ultimate optimization puzzle because the smaller the box, the happier the elves (and Rudolph)!

So grab your coding sleigh, fire up your algorithms, and help Santa pack his trees perfectly for delivery across the globe! The best solution will earn not only the grand prize but also a nod from the big guy himself. Plus, early entries might even receive a special $10,000 reward from Rudolph!

[11/18/2025] Roadmap for Video-Based World Models – This roadmap outlines the progression of video generation systems into full-fledged simulators that combine internal physics-aware world models with visual rendering. It defines four developmental stages toward interactive and stochastic virtual environments.

[11/18/2025] Introducing Gemini 3: Our most intelligent model that helps you bring any idea to life – Today, Google is taking another big step on the path toward AGI and releasing Gemini 3, its most intelligent model that helps you bring any idea to life. Gemini 3 is state-of-the-art in reasoning, the best model in the world for multimodal understanding, and our most powerful agentic and vibe-coding model yet, delivering richer visualizations and deeper interactivity.

Google is releasing Gemini 3 Pro in preview and making it available across a suite of Google products so you can use it in your daily life to learn, build and plan anything. This is just the start of the Gemini 3 era. As of today, Gemini 3 starts rolling out:

  • For enterprises in Vertex AI and Gemini Enterprise
  • For developers in the Gemini API in Google AI Studio, Gemini CLI, our new agentic development platform, Google Antigravity; and more
  • For everyone in the Gemini app and for Google AI Pro and Ultra subscribers in AI Mode in Search
  • In third-party platforms like Cursor, GitHub, JetBrains, Manus, Replit and more

[11/18/2025] PEGATRON Strengthens AI Infrastructure Collaboration with Together AI and 5C for NVIDIA GB300 NVL72 and NVIDIA HGX B200 Liquid-Cooled Rack Deployment in U.S. Data Centers – PEGATRON announced a strategic collaboration with Together AI and 5C to deliver large-scale AI infrastructure across U.S. data centers, powered by NVIDIA GB300 NVL72 and NVIDIA HGX B200 liquid-cooled racks. This partnership highlights the company’s combined expertise in delivering scalable, production-ready, high-performance AI infrastructure, enabling faster, localized service for the U.S. market. Together, the partners are accelerating the rise of AI factories—built for scale, speed, and next-gen AI innovation.

[11/18/2025] Is the Pliops × Zilliz Collaboration Compelling for the AI Industry? – The emerging collaboration between Pliops and Zilliz is well-timed and strategically significant in the current GenAI landscape. As enterprises increasingly adopt Retrieval-Augmented Generation (RAG) to support search, analytics, and large-context applications, they are confronting the growing challenge of scaling vector search to billions of embeddings without prohibitive costs. RAG is moving from an experimental add-on to a core part of enterprise AI architecture, yet the infrastructure needed to support it is straining at the limits of conventional memory, compute, and storage systems.

[11/17/2025] The 1 Billion Token Challenge: Finding the Perfect Pre-training Mix – Most modern language models require massive computational resources and months of training time. This post discusses how to achieve the optimal recipe for creating efficient pre-training data sets to achieve high performance while using dramatically less data. It uncovers critical insights about dataset mixing strategy, failure modes in curriculum testing, and the “goldilocks zone” for synthetic content. The resulting mixture consistently outperforms complex curriculum strategies and avoids catastrophic failures while maintaining excellent generalization.

[11/17/2025] OpenAI-backed FutureHouse launches Kosmos AI Scientist – FutureHouse (which is transitioning to commercial spinout Edison Scientific) released an AI scientist that reads 1,500 papers and runs 42,000 lines of analysis code per run. Beta users estimate that a single Kosmos run accomplishes what would take them six months, with 79.4% of conclusions accurate, though the system sometimes pursues statistically significant but scientifically irrelevant findings. It’s already made seven discoveries—three reproduced unpublished findings and four contributed novel insights, including one related to Alzheimer’s that was validated in human tissue. Kosmos is launching at $200 per run with a generous free tier for academics.

[11/17/2025] Yann LeCun is leaving Meta. And he thinks LLMs are a dead end – One of the founding fathers of modern AI is quietly preparing his next move. And the reason he is leaving might reshape the entire industry. Yann LeCun has spent decades betting on ideas the world thought were impossible. Forty years ago, almost nobody believed neural networks would work at scale. He backed them anyway. He was right. Today he is making another bet. His view is simple: LLMs are impressive. But they are not the future. Not if the goal is real intelligence. Not if the ambition is AGI.

[11/17/2025] Nathan Lambert on open models out-pacing closed labs faster – He highlights three insights: faster open-model releases, benchmark supremacy ahead of real-world user feedback, and the platform shift where Chinese labs challenge U.S. dominance in open models.

[11/17/2025] Andrej Karpathy explains why verifiable work disappears first to AI – The essay reframes AI’s impact by focusing on verifiability. It describes how neural networks improve through practice, why structured tasks advance rapidly, and how this predicts the jagged, uneven frontier of automation.

[11/17/2025] OpenAI’s guide to unlocking GPT-5.1’s full prompting potential – It emphasises that while GPT-5.1 works out of the box, iterative prompt refinement delivers major reliability gains. You’ll learn structure, tool use, and experiment mindset to leverage the model fully.

[11/17/2025] SC25: The International Conference for High Performance Computing, Networking, Storage, and Analysis – St. Louis is the place to be this week (Nov. 16-21) as the high performance computing community convenes for an exhilarating week of sessions, speakers, and networking at its finest. SC is an unparalleled mix of thousands of scientists, engineers, researchers, educators, programmers, and developers who intermingle to learn, share, and grow.

[11/14/2025] AI-orchestrated Cyberattack Detailed by Anthropic – Anthropic disclosed the first documented AI-driven cyber espionage campaign, carried out largely autonomously using its Claude Code tool. The attackers, linked to a Chinese state-sponsored group, targeted 30+ entities worldwide, demonstrating the growing threat of agentic AI systems in cybersecurity.

[11/14/2025] AI startup Cursor raises $2.3 billion funding round at $29.3 billion valuation – Cursor has closed a $2.3 billion funding round at a $29.3 billion post-money valuation. The tool has crossed $1 billion in annualized revenue and swelled to more than 300 employees since it first launched in 2023. Coding tools have proved to be a lucrative AI use case. OpenAI approached Anysphere to potentially purchase Cursor earlier this year, but the deal failed to gain traction.

[11/14/2025] NVIDIA Blackwell Architecture Sweeps MLPerf Training v5.1 Benchmarks – The NVIDIA Blackwell architecture powered the fastest time to train across every MLPerf Training v5.1 benchmark, marking a clean sweep in the latest round of results. As developers experiment with new architectures, and models continue to grow in size, more training compute is essential. Meeting this need for delivered compute requires innovation across every layer of the AI stack—from chips and systems to software—advancing performance at an unprecedented pace.

[11/14/2025] Salesforce AI Research Introduces eVerse, a Simulation Environment for Enterprise-Ready Voice and Text Agents – “Jagged intelligence” — a phenomenon where AI excels at complex tasks but stumbles on simple ones — creates unacceptable business risk. eVerse directly addresses this reliability crisis, building the trust required to make the Agentic Enterprise possible — a business where human employees and AI agents can finally work together in a seamless, collaborative ecosystem. eVerse builds trust in AI outputs through a continuous loop of synthesizing data, measuring performance, and optimizing agent behavior.

[11/13/2025] The French AI lab pleias released a tiny reasoning SLM they call Baguettotron – The 321 million parameter model beats much larger models on reasoning benchmarks. They built it on just 16 H100s with a small team. Baguettotron has an unusual architecture: 80 layers, which is extremely deep for this parameter count (hence the name). They trained it on just 200 billion tokens from a synthetic reasoning dataset they created (with Wikipedia as the anchor) called SYNTH. Their tokenizer is optimized for European languages, and everything is open source.
What’s interesting is the synthetic data angle. It turns out that training on high quality synthetic data is much more efficient than using scraped data and prevents the previously feared model collapse – and it also mitigates the legal concerns around fair use of public, scraped data.
If this holds up at larger scales, it changes the training economics a lot. Smaller teams can compete with frontier labs. You don’t need to scrape the entire internet. You don’t need thousands of GPUs. Training costs drop. Your models can be deployed in more places. Etc.

[11/13/2025] Baidu just dropped an open-source multimodal AI that it claims beats GPT-5 and Gemini – ERNIE-4.5-VL-28B-A3B-Thinking is a 28-billion-parameter multimodal AI model that only activates 3 billion parameters during operation. It has a sophisticated routing architecture that allows it to match or exceed the performance of much larger competing systems on several categories of tasks. The model’s most distinctive feature is ‘Thinking with Images’, a capability that allows the AI to dynamically zoom in and out of images to examine details. The feature dramatically elevates the model’s ability to handle long-tail visual knowledge. The model has been released under the permissive Apache 2.0 license, which allows unrestricted commercial use.

[11/13/2025] Fei-Fei Li’s World Labs speeds up the world model race with Marble, its first commercial product – Marble turns text prompts, photos, videos, or 3D layouts into persistent, editable, and downloadable 3D environments, available now with a free tier. This is the company’s first release after raising $230 million and positions the company as the leading “world model” provider for gaming, VFX, VR, and robotics customers.

[11/13/2025] Synthetic Socratic Debates: Examining Persona Effects on Moral Decision and Persuasion Dynamics – As large language models (LLMs) are increasingly used in morally sensitive domains, it is crucial to understand how persona traits affect their moral reasoning and persuasive behavior. This paper presents the first large-scale study of multi-dimensional persona effects in AI-AI debates over real-world moral dilemmas. Using a 6-dimensional persona space (age, gender, country, social class, ideology, and personality), the researchers simulate structured debates between AI agents over 131 relationship-based cases. The results show that personas affect initial moral stances and debate outcomes, with political ideology and personality traits exerting the strongest influence. Persuasive success varies across traits, with liberal and open personalities reaching higher consensus. While logit-based confidence grows during debates, emotional and credibility-based appeals diminish, indicating more tempered argumentation over time. These trends mirror findings from psychology and cultural studies, reinforcing the need for persona-aware evaluation frameworks for AI moral reasoning.

[11/13/2025] Best LLM Observability Tools of 2025: Top Platforms & Features – Learn about the top LLM observability tools of 2025, including Opik, Langfuse, and Datadog, to monitor, evaluate, and optimize model performance.

[11/12/2025] GPT-5.1: A smarter, more conversational ChatGPT – OpenAI is upgrading GPT‑5 while making it easier to customize ChatGPT. Starting to roll out today to everyone, beginning with paid users: GPT‑5.1 Instant: our most-used model, now warmer, more intelligent, and better at following your instructions; GPT‑5.1 Thinking: our advanced reasoning model, now easier to understand and faster on simple tasks, more persistent on complex ones.

[11/12/2025] Satya Nadella – Inside the world’s biggest AI hyperscaler – As part of this interview, Microsoft CEO Satya Nadella gave Dwarkesh Patel and Dylan Patel (founder of SemiAnalysis) an exclusive first-look at their brand-new Fairwater 2 datacenter. Microsoft is building multiple Fairwaters, each of which has hundreds of thousands of GB200s & GB300s. Between all these interconnected buildings, they’ll have over 2 GW of total capacity. Just to give a frame of reference, even a single one of these Fairwater buildings is more powerful than any other AI datacenter that currently exists. Satya then answered a bunch of questions about how Microsoft is preparing for AGI across all layers of the stack.

[11/12/2025] AMD Financial Analyst Day – AMD rallies after CEO Lisa Su says AI business to grow more than 60% a year for 3 to 5 years. Shares of Advanced Micro DevicesAMD $259.40 (9.00%) are up around 5.3% as of 5:05 a.m. ET, after CEO Dr. Lisa Su unveiled optimistic medium-term revenue guidance at the company’s Financial Analyst Day yesterday. Over the next three to five years, her outlook is for compounded annual revenue growth of over 35% for AMD as a whole, with growth north of 60% for its data center business.

[11/12/2025] MLPerf Training v5.1 – In this round, MLCommons published 185 performance results from 20 organizations, including 3 new submitters: DataCrunch, the University of Florida, and Wiwynn. The organization introduced two new training benchmarks, and the results highlighted an increase in system diversity and also a greater emphasis on larger multi-node systems. We saw performance gains well over 2X on key generative AI workloads; collectively the results in this round all point to a substantial improvement in capabilities for the whole AI community.

As we all know, the pace of innovation in AI is incredibly rapid, and in this release MLCommons updated two benchmarks to better represent state-of-the-art technology solutions. In particular, Llama 3.1 8B for text summarization was introduced, replacing the venerable BERT reference model and Flux.1 replaces Stable Diffusion v2 to offer a more modern approach for text-to-image models.

The published results show an increasing emphasis on generative AI (genAI) tasks, consistent with trends in the industry. Looking at the Llama 2 70B LoRA benchmark for finetuning, the number of results increased by 24% compared to prior rounds. Similarly, if we compare the new Llama 3.1 8B summarization benchmark, the number of results increased by 15% over its predecessor (BERT), which is quite impressive for a brand new benchmark.

The latest round of results is incredibly diverse with a total of 65 distinct systems, showcasing 12 hardware accelerators and a wide range of software stacks. Notably, almost half of these were multi-node setups, representing an 86% increase compared to the v4.1 round held a year earlier. These multi-node submissions leveraged various network architectures, with many relying on custom-built solutions.

[11/12/2025] A tech leader’s guide to AI on Kubernetes – Google Cloud has invested years into developing a robust managed container platform built to handle the unique demands of AI and other big leaps in technology evolution. Google Kubernetes Engine (GKE) offers advanced orchestration and management while reducing the operational challenges that come with scaling and deploying new AI solutions.

Download this ebook to fuel AI innovation with GKE and:

  • Conquer AI challenges with confidence: Balance costs, performance, and security, turning complexities into opportunities among the rising use of gen AI and large language models.
  • Unleash AI at scale: Develop, deploy, and manage AI solutions at an unprecedented scale, unlocking competitive advantages and accelerating innovation.
  • Get on the path to profitability: Achieve significant cost savings with faster training and serving times, boosting your ROI while simplifying operations.

[11/11/2025] Datometry Joins Snowflake – Snowflake and Datometry have entered into a definitive agreement with the aim of taking the Datometry technology to the next level.

[11/11/2025] New leaked image samples by Nano Banana 2 – Nano Banana 2 access was briefly leaked, and image samples were widely circulated across social media. The model appears capable of remastering low-quality images and following detailed text prompts. There is still no confirmation of the release date, but a launch is likely to be imminent. The leaked samples are available in the article.

[11/11/2025] The 4.5 trillion dollar elephant in the room – NVIDIA, currently valued at $4.5 trillion, is the most valuable company on Earth and the dominant supplier of AI chips. A pattern of fear around the company is impacting policy discussion. There are claims that the major AI companies disagree with NVIDIA’s policy stances, but are unwilling to speak out because NVIDIA has the power to undermine their business. This article looks at the evidence of such claims.

[11/11/2025] AI Diffusion Report: Where AI Is Most Used, Developed, and Built – AI has reached 1.2 billion users in under three years, marking the fastest technology adoption in human history. However, the Global North’s adoption rate is nearly double that of the Global South. The next 1–2 billion users may come easily, but adoption is expected to plateau afterward because the remaining ~4 billion people lack electricity or internet access. Low-resource languages create an additional barrier: half the web is in English, even though only 5% of people speak it natively.

[11/11/2025] You can now test upcoming GPT-5.1 on OpenRouter – OpenAI’s latest code updates reference a general availability date of November 24 for the GPT-5.1 family for Enterprise accounts. Plus and Pro users may gain access to the models even earlier, possibly within the coming week. GPT-5.1 Thinking may currently be testing under the name Polaris Alpha on platforms such as OpenRouter. Polaris Alpha has a 256k context window and quite good performance.

[11/10/2025] For $300M, Megaport buys Latitude.sh to expand data center infrastructure offering – The data center networking provider Megaport announced today that it had bought Latitude.sh, a server hardware company, for $300 million. The deal will allow Megaport to directly own the compute infrastructure and a sizable portion of the global inference AI infrastructure market.

Latitude.sh provides high-performance CPU and GPU bare metal infrastructure globally while Megaport provides the interconnection capacity for data centers to connect with each other and end users. Megaport serves over 1000 data centers in 125 cities worldwide. The addition will allow Megaport to better serve the burgeoning inference AI demand.

[11/10/2025] The state of AI in 2025: Agents, innovation, and transformation – McKinsey released its State of AI 2025 survey of nearly 2K organizations, revealing that while almost every company now uses AI, most are stuck in pilots, with only a fraction achieving enterprise-wide impact or scaling agents.

Key details:

  • The survey found that 88% of companies now use AI somewhere, but most of them are in experimentation or pilot phases, with just 33% actually scaling it.
  • While 39% reported EBIT impact from AI, just 6% achieved an impact of 5% or more, largely by redesigning workflows and using it to drive innovation.
  • 62% are working with AI agents, but adoption is early, with 39% experimenting and just 23% scaling them, mostly in IT and knowledge management.
  • About 32% of companies expect workforce reductions of 3% or more next year, while 13% expect increases. Larger firms are more likely to predict cuts.

The key lesson comes from the high performers — the few seeing real bottom-line impact from AI. Their success shows that the real value of AI comes not from efficiency gains, but from redesigning workflows, scaling across functions, and using it to fuel growth and innovation.

[11/10/2025] If your model performs poorly, don’t just tune hyperparameters. Check:

  • Label leakage
  • Temporal splits
  • Variable encoding
  • Metric mismatch

Most errors stem from setup, not the algorithm itself.

[11/10/2025] Inception raises $50 million to build diffusion models for code and text – With so much money flooding into AI startups, it’s a good time to be an AI researcher with an idea to test out. And if the idea is novel enough, it might be easier to get the resources you need as an independent company instead of inside one of the big labs. That’s the story of Inception, a startup developing diffusion-based AI models that just raised $50 million in seed funding. The round was led by Menlo Ventures, with participation from Mayfield, Innovation Endeavors, Microsoft’s M12 fund, Snowflake Ventures, Databricks Investment, and NVIDIA’s venture arm NVentures. Andrew Ng and Andrej Karpathy provided additional angel funding.

[11/10/2025] OpenAI Outlines How to Handle the Next Phase of AI Progress – OpenAI published an update on current AI trajectories and the policies needed to manage them. The post combines internal data on scaling with recommendations for governments and research institutions to prepare for the next wave of capability jumps.

Key details:

  • OpenAI estimates that AI “intelligence per dollar” is improving roughly 40× per year, faster than past hardware scaling trends.
  • The team expects smaller breakthroughs in 2026 followed by a larger step-change in 2028 or later, as research cycles mature.
  • The post calls for international coordination on model safety testing, reporting, and evaluation, similar to aviation or cybersecurity standards.
  • It recommends public funding for red-team exercises, AI safety research, and stress-testing to improve system resilience.
  • OpenAI also supports monitoring “compute concentration” to track who trains large models and at what scale.

The report frames AI progress as predictable in trend but uncertain in outcome, arguing that preparedness, not prediction, should guide policy and research investment.

[11/10/2025] Google launches Nested Learning, a new brain-inspired method to fix AI’s catastrophic forgetting – For years, deep learning has relied on a simple assumption: a model is one big optimization problem that learns from data all at once. But that simplicity has a price, catastrophic forgetting. Train a model on new data, and it often forgets what it already knew. Google Research wants to fix that with a new paradigm called Nested Learning, unveiled with a proof-of-concept architecture named Hope.

The problem: memory that doesn’t last – Machine learning models struggle to retain knowledge over time. Even large language models, with billions of parameters, lose context when fine-tuned or retrained. This is the same issue that prevents continual learning, the ability to learn new skills without erasing old ones.

The insight: models are more like brains than we thought – Nested Learning reimagines a model not as one giant learner but as a collection of smaller learners nested inside each other. Each one runs its own optimization loop and updates at its own speed, just like different brain regions processing information at distinct rhythms.

The breakthrough: self-modifying and context-aware learning – Google’s Hope model uses Nested Learning to structure memory across multiple timescales. Instead of treating architecture and training as separate, it unifies them into one system of layered optimization problems. This means the model can decide which parts to update quickly and which to keep stable.

Hope outperforms Titans, Samba, and Transformers on language modeling and reasoning tasks. It achieves lower perplexity, higher accuracy, and handles long-context reasoning better, solving “Needle-in-a-Haystack” tasks with more stability and precision. You can simulate nested optimization today in PyTorch or JAX by assigning different update rates to model components or optimizers. This lets you fine-tune large models while preserving old knowledge, a practical first step toward continual learning.

[11/10/2025] Google launches File Search tool in the Gemini API – Building production-ready RAG (Retrieval Augmented Generation) applications is traditionally a heavy engineering lift. Today, Google is launching File Search for the Gemini API to dramatically simplify this process. File Search, built directly into the Gemini API, is a fully managed RAG system that abstracts away the retrieval pipeline—automatically handling file storage, chunking strategies, and dynamic content injection so developers can focus on building rather than managing infrastructure.

Why it matters:

  • Zero-hassle RAG: It’s integrated directly into the Gemini API, making it easy to ground Gemini models with proprietary data (PDFs, JSON, common code formats, etc.).
  • Disruptive Pricing: Storage and query-time embedding generation are free. Developers only pay for initial indexing at a fixed rate of $0.15 per 1 million tokens.
  • Built-in Trust: Responses include automatic citations, pointing specifically to the source data used to generate answers.

Early adopters like Phaser Studio are already using it to power their AI game generation platform, reducing tasks that once took hours of manual cross-referencing to under 2 seconds. Adding to the momentum of new features dropped almost daily throughout the last month, the excitement around new features across Google’s developer platforms has continued to grow. Logan Kilpatrick just announced 2.1M users use Google AI Studio to vibe code hundreds of thousands of apps made every day.

[11/7/2025] Introducing Nested Learning: A new ML paradigm for continual learning – Google has introduced Nested Learning, a new approach to machine learning that views models as a set of smaller, nested optimization problems, each with its own internal workflow, in order to mitigate or even completely avoid the issue of “catastrophic forgetting,” where learning new tasks sacrifices proficiency on old tasks.

[11/7/2025] Microsoft establishes new Superintelligence Team – Microsoft AI CEO Mustafa Suleyman announced the MAI Superintelligence Team, a research division dedicated to building advanced systems that solve specific problems in areas like medicine and energy over open-ended AGI.

Key details:

  • Suleyman emphasized building “Humanist Superintelligence”, prioritizing AI that “always works for, in service of, people and humanity more generally.”
  • The team will focus on narrow, high-impact societal challenges, highlighting AI learning companions, medical superintelligence, and clean energy advances.
  • Suleyman’s Inflection co-founder Karen Simonyan is serving as chief scientist alongside poached researchers from DeepMind, OAI, and Anthropic.
  • The announcement comes in the wake of Microsoft’s new arrangement with OAI, allowing both to pursue superintelligence independently of each other.

We’ve seen all the major AI labs forge their own unique vibe over the last few years, and it feels like Suleyman has finally crafted one for Microsoft’s AGI efforts — forging a “humanist” identity and unique direction that felt missing for much of the tech giant’s initial OAI arrangement.

[11/7/2025] Google deploys new Axion CPUs and seventh-gen Ironwood TPU — training and inferencing pods beat NVIDIA GB300 and shape “AI Hypercomputer” model – Google Cloud has introduced new AI-oriented instances powered by its own Axion CPUs and Ironwood TPUs. The new instances are aimed at both training and low-latency inference. Servers running these chips come equipped with Google’s custom Titanium-branded controllers, which offload network, security, and I/O storage processing, enabling better management and higher performance. More details about Google’s in-house designed processors are available in the article.

[11/7/2025] Leaving Meta and PyTorch – Soumith Chintala has left Meta after working at the company for 11 years. Chintala spent almost eight years leading PyTorch, taking it from nothing to more than 90% adoption in AI. He plans to move to something small and new. PyTorch has been left in capable hands, and Chintala is confident that its values will stay intact, but things may change in flavor because he will no longer impose his own taste from the top.

[11/7/2025] Accumulating Context Changes The Beliefs of Language Models – Extended conversations can reshape how AI models answer the same question. After 10 rounds of debate about moral dilemmas, GPT-5 reversed its position 54.7% of the time when asked again. When models changed their stated beliefs, they also chose different tools to accomplish tasks.

[11/7/2025] Kimi K2 Thinking – Chinese lab Moonshot AI released an open-source reasoning model that matches or exceeds GPT-5 and Claude Sonnet 4.5 on some benchmarks, including a 44.9% on Humanity’s Last Exam and 71.3% on SWE-Bench Verified.

[11/7/2025] Forecasts of AI & Economic Growth – Most economists forecast 0.1-1.5% annual excess growth through 2035, while AI insiders predict 3-30%. Most economists, though, assume that today’s ChatGPT is essentially the ceiling of AI progress and that visions of a virtual AI coworker won’t really materialize, but how is it possible that cheap, PhD-level intelligence wouldn’t boost productivity? The forecasts imply that either AGI will be far less general than claimed or that adoption will be dramatically slower than historical technology diffusion. AI’s 2024 economic footprint appears to be around 0.5% of GDP.

[11/7/2025] OpenAI readies GPT-5.1 Thinking model ahead of Gemini 3 Pro – GPT-5.1 Thinking has been spotted in a backend component responsible for powering advanced reasoning processes within ChatGPT, indicating a GPT-5.1 release is close. Rumors point to a lineup of three models (Mini, Thinking, and a possible Codex-focused upgrade), each designed to fit different user needs and compute constraints. OpenAI is likely aiming to synchronize the release with Google’s upcoming Gemini 3 Pro launch.

[11/7/2025] Mathematical exploration and discovery at scale – Google DeepMind has been collaborating with Terence Tao and Javier Gómez-Serrano to use our AI agents (AlphaEvolve, AlphaProof, & Gemini Deep Think) for advancing Maths research. They find that AlphaEvolve can help discover new results across a range of problems. As a compelling example, they used AlphaEvolve to discover a new construction for the finite field Kakeya conjecture; Gemini Deep Think then proved it correct and AlphaProof formalized that proof in Lean. AI-powered Maths research in action!

[11/6/2025] Beyond Standard LLMs – The largest and most capable open-weight large language models today are autoregressive decoder-style transformers. Some alternatives have popped up in recent years. While some are geared toward better efficiency, others are aimed at improving model performance. This article by LLM industry luminary Sebastian Raschka gives an overview of alternative models that exist, from text diffusion models to the most recent linear attention hybrid architectures.

[11/6/2025] How AI Scaling Disrupts Economic Theory – Harvard computer scientist Boaz Barak published an essay on how rapid AI progress might disrupt traditional economic models. He uses data from AI capability scaling to test whether current theories of productivity and GDP growth still hold under exponential improvement.

Key takeaways from the essay:

  • METR data shows top LLMs double their task-solving horizon every six months.
  • Historical U.S. GDP per-capita grows ~2% annually, doubling every 35 years.
  • Standard models can’t handle a labor source that scales both in quantity and quality exponentially.
  • If AI creates virtual workers whose number and skill both rise, GDP could increase sharply.
  • Growth remains capped by the share of work that resists automation even if AI productivity becomes infinite.
  • Real-world frictions like coordination, physical work, and diffusion, limit how fast theory turns into output.

Deterministic AI scaling challenges economic assumptions built for gradual change, forcing a rethink of what “growth” means in an automated economy.

[11/6/2025] LMArena launches Arena Expert to evaluate actual expert-level prompts – It introduces a new Expert leaderboard, and also maps all prompts into 23 occupational fields. This shifts LLM evaluation beyond static benchmarks and into expert-grade task difficulty. It lets you see if a model can solve real advanced problems in domains like software engineering, medicine, and law.

[11/6/2025] Tencent and Tsinghua show LLMs can skip next-token prediction – Tencent and Tsinghua propose CALM, which stops predicting one token at a time. It compresses chunks of tokens into continuous vectors, then predicts the next vector. This cuts generation steps and shows you can scale language models without discrete token limits.

[11/6/2025] Efficient Reasoning for Specialized AI Agents With NVIDIA Nemotron – AI agents rely on diverse models to understand visual inputs, retrieve information, generate content, apply guardrails, and reason through complex tasks.

At GTC Washington, D.C., NVIDIA unveiled the newest additions to the NVIDIA Nemotron™ family of world-class open models and datasets, including video, retrieval, and content-safety capabilities, enabling developers to build and deploy specialized AI agents.

[11/6/2025] Solve Large Linear Programming Problems With the NVIDIA cuOpt Barrier Method Solver – Solving linear programming problems is no easy task, since they contain millions of variables and constraints that can slow CPUs to a crawl. They’re solved using three main methods: simplex, primal-dual linear programming (PDLP), and barrier.

The latest release of NVIDIA® cuOpt™ includes a new, award-winning, GPU-accelerated solver for linear programs that uses the barrier method. This new solver complements the existing cuOpt LP solvers, enabling engineers, researchers, and developers to solve large-scale linear programs quickly and accurately.

Explore the new notebook to learn how to install the latest cuOpt release, run the new barrier method, and control important settings like presolve and folding.

[11/6/2025] Introducing OlmoEarth—State-of-the-art foundation models & open infrastructure to tackle the planet’s biggest problems. This week, Ai2 released OlmoEarth—planet-scale foundation models for people, communities, and our planet. In addition to the state-of-the-art models, the company also released OlmoEarth Studio for Earth data, annotation, and fine-tuning, part of our new OlmoEarth Platform.

OlmoEarth is an open, scalable, end-to-end platform leveraging our new state-of-the-art Earth observation foundation models. The platform supports end-to-end workflows for building and serving Earth AI, from raw data collection to annotations and label management, through fine-tuning and deployment. OlmoEarth foundation models are highly efficient and easy to train, enabling seamless customization to support a wide variety of use cases. OlmoEarth’s integrated workflows cut costs and development time to better serve urgent, real-world needs.

[11/6/2025] In AI, speed matters. But accuracy matters more. The companies that win aren’t just those who ship fast—they’re the ones who test assumptions, build trust, and design for change. Fast ≠ reckless.

[11/5/2025] arXiv pushing back against “AI slop” – arXiv, the leading open-access repository for scientific papers, announced it will no longer accept computer science review articles and position papers due to a surge of low-quality, AI-generated submissions. The site said its CS section has been overwhelmed by papers that are “little more than annotated bibliographies,” often produced using large language models.

While arXiv has always required original research, it will now enforce stricter rules, review and opinion papers must show evidence of peer review to be considered. The move aims to curb the flood of AI-generated “slop” and free moderators to focus on substantive research. arXiv warned that other categories could see similar restrictions if trends continue, reflecting broader concerns in academia about the rise of AI-created junk science and declining quality control in research publishing.

[11/5/2025] Why voice is the next AI interface – ElevenLabs CEO Mati Staniszewski says voice will be the next major human-computer interface, driving both creative and enterprise adoption. ElevenLabs operates as a remote-first company, hiring globally and maintaining a flat structure without traditional titles to promote speed and accountability.

Its Voice Marketplace, which has paid creators over $10 million, evolved from a creator-focused platform into an enterprise solution, despite challenges like complex licensing negotiations and long sales cycles. Staniszewski also reflected on lessons from early hiring, scaling, and balancing cutting-edge research with fast product launches—all while staying independent and occasionally turning down licensing deals with competitors.

[11/5/2025] Deepnote, the AI Notebook, Goes Open Source – Deepnote, now open source under Apache 2.0, lets you analyze data, build dashboards, and collaborate with AI, all in one place.

Built as the successor to Jupyter, it stays fully compatible while adding a built-in AI agent, 100+ data integrations, and a human-readable .deepnote YAML format for clean version control.

It allows you to:

  • Run code, SQL, and visualizations together
  • Connect to data sources instantly
  • Collaborate in real time with AI assistance
  • Keep notebooks reactive, updates flow automatically

[11/5/2025] Gemini API update makes AI agents more reliable – Building reliable AI applications just got significantly easier. Today, Google launched major improvements to Structured Outputs in the Gemini API, guaranteeing AI responses strictly adhere to industry-standard data schemas, including JSON, to remove the unpredictability that often breaks production applications.

For developers, this means spending less time fixing unexpected AI outputs and more time building reliable production systems. Direct JSON Schema support allows popular data validation libraries to be used out-of-the-box with the API.

Here’s what it unlocks:

  • Seamless Agent Communication: It creates a common language for multiple AI agents. One agent’s output becomes another’s perfect input, enabling complex systems to collaborate without translation layers.
  • Direct Database Population: Developers can now format data to be inserted directly into tables without intermediate cleaning steps.
  • Reliable Data Extraction: It ensures precise details can be pulled from unstructured sources, like emails, with guaranteed accuracy.

[11/5/2025] Google Cloud announces Vertex AI Agent Builder updates for production-scale agents – Google Cloud is announcing new capabilities for Vertex AI Agent Builder, focused on solving the developer challenge of moving AI agents from prototype to a scalable, secure production environment.

With these updates, developers can:

  • Build faster with SOTA context management layers as well as one-click deployment in ADK
  • Scale in production with new managed services from the Vertex AI Agent Engine (AE) including new observability and evaluation capabilities
  • Govern agents with confidence with new features including native agent identities and security safeguards

[11/5/2025] Gaining buy-in: A guide to successful AI integration in development – AI-assisted coding is revolutionizing software development. The presentation below moves beyond theory to showcase the practical applications of AI-assisted coding within large enterprise teams, highlighting the specific ways developers are utilizing Gemini Code Assist – Google’s AI coding assistant – to streamline workflows. You will learn from this session how to successfully gain buy-in from both management and engineering teams by demonstrating the tool’s value and addressing potential concerns. Learn when and how to integrate new AI technologies while prioritizing security, responsible usage, and alignment with existing development processes.

[11/4/2025] Training LLMs end to end is hard. Very excited to share a new blog (book?), “The Smol Training Playbook: The Secrets to Building World-Class LLMs,” that cover the full pipeline: pre-training, post-training and infra. 200+ pages of what worked, what didn’t, and how to make it run reliably.

[11/4/2025] Inside NeurIPS 2025: The Year’s AI Research, Mapped – NeurIPS 2025 (San Diego Convention Center, Dec. 2-7) papers are out—and it’s a lot to take in. This visualization lets you explore the entire research landscape interactively, with clusters, summaries, and @cohere LLM-generated explanations that make the field easier to grasp.

[11/4/2025] Notes from reading the Amp manual – Amp is a coding agent with support for MCP, the ability to set project, folder, and global context, a permissions system, headless mode, and more. It stores all threads on the server so users can resume a thread anytime. Amp picks the right model for each task, but it can be nudged to use more expensive models. It is pay-per-token and not cheap, but it is worth it because it saves time.

[11/4/2025] What’s up with Anthropic predicting AGI by early 2027? – To get Nobel-Prize-winning AIs by early 2027, AIs must fully automate researching engineering by December 2026 and go from completing hour-long tasks today to completing two-week-long tasks in just the next 8 months. OpenAI is only slightly more conservative than Anthropic, estimating automated AI research by March 2028.

[11/4/2025] SALESFORCE REPORT: 84% of tech leaders say their data can’t support AI ambitions – Salesforce today released its State of Data & Analytics Report revealing the massive data overhaul required for enterprises to successfully execute their AI strategies.

While 63% of business leaders describe their organizations as data-driven, the same amount of technical leaders say they struggle to drive business priorities with data – even though 76% say they are under growing pressure to drive business value with data. The core problem? Incomplete, out-of-date, or poor-quality data that prevents companies from being truly data-driven.

To close the gap, technical leaders are embracing trusted, unified data, adopting stronger governance, and implementing zero-copy architectures to solve for data silos. In fact, companies using zero-copy are 34% more likely to succeed with AI initiatives.

[11/3/2025] Nobel Laureate Hinton on AI – A year after winning the Nobel Prize for his work related to machine learning, Geoffrey Hinton, the “Godfather of AI,” is sounding the alarm. He says that artificial intelligence could one day outsmart and overpower its creators. Hinton warns that tech giants are moving too fast, that global cooperation is needed, and that we may need a “Chernobyl moment” to take the danger seriously. From mass job losses to loss of control, Hinton says humanity’s survival depends on how we act now.

[11/3/2025] Ghosts in the Codex Machine – After weeks of user reports claiming Codex had gotten worse, echoing a perennial concern that frontier labs secretly downgrade their models, OpenAI assembled a squad to investigate every possible culprit and publish their findings. The team found no smoking gun, but instead uncovered a catalog of small bugs: older hardware subtly dragging down performance, a constrained sampling bug that made the model switch to Korean mid-sentence, and context compaction that quietly degraded accuracy in long conversations.

[11/3/2025] From Ilya’s deposition – Ilya Sutskever plotted for over a year with Mira Murati to remove Sam Altman as OpenAI CEO. Dario Amondei wanted Greg Brockman fired and himself in charge of all research. Murati told Sutskever that Altman had pitted her against Daniela Amondei. Sutskever wrote a 52-page-long memo to get Altman fired and also a separate document on Brockman. OpenAI is paying Sutskever’s legal bills.

[11/3/2025] Gen AI fast-tracks into the enterprise – Generative AI has become mainstream in the enterprise, with widespread workplace adoption and growing familiarity, though depth and maturity still vary across organizations. Leaders are increasingly focused on measuring value, ROI, and accountability, as early returns build conviction and optimism about AI’s long-term potential. At the same time, human capital remains the key enabler—executive engagement has surged, with 61% of companies now employing Chief AI Officers—but training and capability development still lag behind ambitions, highlighting a gap between enthusiasm and readiness.

[11/3/2025] Sam and Satya talk $3 trillion AI buildout – Brad Gerstner interviews Satya Nadella (Microsoft) and Sam Altman (OpenAI) about the $3 trillion AI buildout reshaping global technology and the economy. They explore how the Microsoft–OpenAI partnership enabled massive cloud scaling and transformed both companies’ strategies. Nadella discusses Microsoft’s Azure and Copilot expansion to meet surging AI demand, while Altman reflects on AI’s power, human ambition, and future milestones toward AGI.

Topics include OpenAI’s nonprofit structure, compute constraints, regulation, AI hardware, and the economic ripple effects on productivity and jobs. The conversation highlights a vision of AI as the new industrial revolution, driving both digital and physical reindustrialization in the U.S. by 2026 and beyond.

[11/3/2025] Hottest new job in AI – About a year ago, the hottest AI job was prompt engineer, specialists who fine-tuned chatbot responses through clever wording. Now, the new must-have role is the forward-deployed engineer (FDE). Companies like OpenAI, Anthropic, and Cohere are rapidly hiring these hybrid professionals who can both code and collaborate directly with customers to tailor AI models for real-world use.

Job listings for FDEs have surged over 800% in 2025, according to Indeed, as enterprises from manufacturing to healthcare seek help integrating AI into their operations. The term “forward-deployed” originates from the military and was popularized by Palantir, which sends embedded engineers into client environments to co-create solutions.

OpenAI plans to grow its FDE team to around 50 engineers, helping clients like John Deere use AI to reduce chemical spraying by up to 70%. These embedded engineers are becoming essential to translating cutting-edge AI research into practical, revenue-driving products.

[11/3/2025] AWS and OpenAI announce multi-year strategic partnership – The multi-year, strategic partnership empowers OpenAI with immediate and increasing access to AWS’s world-class infrastructure for their advanced AI workloads. AWS to provide OpenAI with Amazon EC2 UltraServers, featuring hundreds of thousands of chips, and the ability to scale to tens of millions of CPUs for its advanced generative AI workloads. Representing a $38B commitment, OpenAI will rapidly expand compute capacity while benefitting from the price, performance, scale, and security of AWS.

Posted on 11/21/2025, in AI, Analytics, Big Data, cloud, Data Science, Data Storage, Database, Deep Learning, GenAI, Hardware, Machine Learning, News, Security, Uncategorized and tagged AI, Data Science, GenAI, LLM, Machine Learning. Bookmark the permalink. Leave a comment.