Artificial Intelligence
Explore the latest in machine learning, deep learning, natural language processing, and AI applications
Navigate through content by publication date
Tue, Nov 17
74 items found
How Much Does Saliency Accuracy Matter for Saliency-Guided 360-degree Image Compression?
Saliency-guided coding allocates bits to the regions viewers are most likely to attend, but accurate 360-degree saliency prediction is computationally expensive, an obstacle for low-latency immersive streaming applications such as robot teleoperation. We quantify how much saliency-prediction accuracy actually translates into compression gain. Building on a saliency-guided rate-distortion-optimized (RDO) learned codec, modified so that the training-time latent mask is driven by ground-truth saliency, which yields a single, predictor-agnostic codec, we evaluate saliency predictors spanning two orders of magnitude in compute on both the learned codec and H.265/HEVC. The results show that higher saliency-prediction accuracy yields diminishing compression gains: a lightweight predictor fast enough for live use (52 frames per second) captures 65% of the ground-truth oracle's bit-rate saving, a 3.3x heavier one reaches 74%, and a further 38x increase in compute adds no measurable benefit, on either codec. Therefore, lightweight predictors might also be a practical choice for latency-constrained 360-degree streaming.
Reading or Remembering? Measuring Temporal Contamination in Language Models
Language models are now used to read central-bank text for economic forecasting, but most were trained on text written after the documents they read. A model reading a 2021 FOMC statement may already know that inflation rose in 2022, so part of its apparent skill can be memory of the outcome, a temporal form of test-set contamination. We measure this memory on FOMC statements and minutes from 2000 to 2026 by comparing modern open models with chronologically consistent models, which are trained only on text from before a fixed date. Asked about past GDP outcomes given only a date, models that cannot know the answer still reach an AUC of up to 0.64 from priors on dates and eras, so chance is the wrong baseline. Against these memory-free models on the same questions, most modern models recall which years had strong growth. Membership-inference scores sort documents by era and document type rather than by whether a model saw them, as exact counts in an open pretraining corpus show. In real-time nowcasts, every lookahead premium we estimate on US GDP is smaller than the forecast comparison can detect at macroeconomic sample sizes, and none on GDP or inflation survives correction for multiple testing. Lookahead audits should therefore measure memory directly against memory-free models on the same items.
When Do Histology Foundation Models Need Geometric Test-Time Augmentation?
Test-time augmentation (TTA) aggregates model predictions across multiple transformed views of an input during inference. Because histology images lack a canonical orientation, $D_4$ TTA---pooling across the 8 rotations and reflections of a square---is a standard default in computational pathology. However, the rise of self-supervised histology foundation models (FMs) with inherent rotational invariance calls its continued utility into question. We benchmark 9 histology FMs and 6 general-purpose models across four patch-level and two whole-slide image (WSI) datasets. We find that $D_4$ TTA remains a reliable post hoc improvement, but benefits diminish from $+1.98$ percentage points (pp) for general-purpose backbones to $+0.50$ pp for histology FMs on TCGA-UT, and shrink further for the largest models. To eliminate computational overhead, our experiments demonstrate that entropy routing---applying TTA selectively to uncertain predictions---recovers most of the gain while augmenting only ${\sim}13\%$ of samples. Finally, we show that patch-level TTA improvements do not transfer to slide-level multiple instance learning (MIL). Code: https://github.com/lars-chen/histology-tta
How VLMs Plan What's Next
When a vision-language model describes an image, it names one object at a time and it rarely repeats itself, yet how the model plans this order is far from obvious. We characterize the mechanism a VLM uses to decide where to look before it looks. The mechanism has four main parts. First, binding heads match the relevant text tokens against the image and resolve them into pointers to the matched image region. Second, the same binding heads bind each written answer to the region it came from, and the model turns it into a progress signal: a cursor, the position in the image up to which the list has been written. Third, a select step chooses the nearest unwritten region after the cursor, or stops. Finally, gaze heads execute the chosen pointer and return a concept of what is drawn there, which language heads and the late MLPs turn into a word in the language of the question. We validate each part of this mechanism through causal patching in Qwen3-VL models on a synthetic visual retrieval task, and test our mechanisms against alternative hypotheses. We further show that the mechanism holds across input formats, natural photographs, and languages, and that its binding, relay, and gaze heads appear at every model size from 2B to 32B. This work allows us to investigate how a VLM turns a question and an image into a sequence of answers, layer by layer and token by token. More broadly, it opens a way to study how models plan over complex multi-step tasks with long rollouts. Our project is available at https://plan.baulab.info/
Severe but Harmless: Stress-Testing Shift Severity for Frozen Time-Series Forecasters
Zero-shot time-series foundation models (TSFMs) have become attractive alternatives to task-specific models because they handle diverse downstream tasks, including anomaly detection and forecasting, without per-series training. However, time-series data undergo distribution shifts, and the dissimilarity between two regimes is commonly scored with statistical measures such as Wasserstein-1, where a higher score is read as a more severe shift. In systematic experiments using three established severity measures on nine regime shifts from three standard traffic datasets, we show that these measures do not reliably predict degradation for instance-normalized TSFMs. The regimes rated most severe are not the regimes that degrade the models most, and degradation reaches 29% in the lowest-scoring regime. In most cases, the audit shows no positive relationship between measured severity and model degradation. We provide an equivariance-based stress-testing analysis that explains why severity readings can become decoupled from model degradation.
Mitigating Catastrophic Forgetting with Physics-Indexed Specialist Memory: the Evolutionary Neuroplastic Agent
Continual reinforcement learning requires adaptation to changing conditions while preserving performance under previously encountered dynamics. It was investigated whether specialists generated by evolutionary search can provide a reusable memory when combined with environment detection and failure-driven search. The Evolutionary Neuroplastic Agent (ENA) evolves small neural networks without gradients, archives selected policies by detected dynamic regime, and retrieves them through Bayesian policy reuse. The architecture was evaluated in non-stationary CartPole and HalfCheetah environments. In the 30-replicate CartPole ablation, the complete architecture achieved a mean success rate of 83.2% across the five test environments. Among variants that learned the harder CartPole environments, the complete architecture showed the most consistent retention. A selection ablation also showed that the retrieval rule affected performance. In the evaluated sequences, ENA adapted more slowly than PPO on CartPole and achieved lower returns than SAC on HalfCheetah. The evidence concerns the complete architecture and does not isolate gradient-free optimization as the cause of retention or establish superiority over other continual-learning methods.
Your Browser Can Run PyTorch Models Now. Here’s How WebGPU Makes It Possible
No server, no install, no data leaving the device. What changed, what it costs, and where it still breaks.Continue reading on Towards Deep Learning »
Hugging Face PyBot: 1,000 Agents Built a Chatbot in Python Using Only If-Else Statements
PyBot is 640,000 lines of if-statements that tie a 360 million parameter language model on everyday chat.Continue reading on Towards Deep Learning »
Medical Image Segmentation in 2026: Do Foundation Models Really Beat the Classic U-Net?
From nnU-Net to MedSAM and SAM 3: What recent research reveals about the future of medical image segmentation.Continue reading on Towards Deep Learning »
LLM’lerin Cevabına Nasıl Güvenebiliriz? Structured Output ve Middleware
Bir önceki yazıyı şu soruyla bitirmiştik: Uygulamamız JSON beklerken model açıklama metniyle birlikte geçersiz bir çıktı üretirse ne…Continue reading on Medium »
Gemini Will Now Make Phone Calls for You (US Only)
The November 2026 Pixel Drop adds Gemini calling on your behalf for Pixel 10 and newer. We list what it does, where it works, and what…Continue reading on Medium »
Anthropic Cut Junior Tech Hires From 27% to 9%. Copy That, and You Run Out of Seniors.
Australia’s central bank now sees young workers hit first in AI-exposed jobs. Companies that stop hiring juniors because AI does junior…Continue reading on Towards AI »
Your Browser Can Run PyTorch Models Now. Here’s How WebGPU Makes It Possible
No server, no install, no data leaving the device. What changed, what it costs, and where it still breaks.Continue reading on Towards Deep Learning »
Hugging Face PyBot: 1,000 Agents Built a Chatbot in Python Using Only If-Else Statements
PyBot is 640,000 lines of if-statements that tie a 360 million parameter language model on everyday chat.Continue reading on Towards Deep Learning »
Medical Image Segmentation in 2026: Do Foundation Models Really Beat the Classic U-Net?
From nnU-Net to MedSAM and SAM 3: What recent research reveals about the future of medical image segmentation.Continue reading on Towards Deep Learning »
Boyuan Capital Closes RMB 1.1B Fund II to Back Industrial AI
Boyuan Capital announced the final close of Fund II on November 17, 2026, at approximately RMB 1.1 billion. The fund will focus on industrial artificial intelligence and embodied intelligence, alongside business creation within Bosch’s industrial value chain and overseas expansion by Chinese advanced-manufacturing companies. Boyuan describes itself as the Bosch Group’s only market-oriented investment platform. Its role is to connect Chinese technological innovation with global industrial…
98 percent of scientists expect AI to accelerate scientific discoveries, OpenAI study finds
Artificial intelligence could significantly shorten the time needed for scientific breakthroughs, with 98 per cent of surveyed scientists expecting AI to accelerate discoveries in their fields, according to a new study by OpenAI, highlighting the growing role of tools such as ChatGPT in research.
How Check Point speeds up vulnerability research with Daybreak | OpenAI for Cybersecurity
Check Point CTO Jonathan Zanger explains how his team uses OpenAI Daybreak to analyze vulnerabilities and turn findings into protections for customers. A vulnerability research workflow that once took hours can now be completed in minutes, according to Zanger. Read the full customer story: https://openai.com/index/check-point-daybreak/
How Check Point speeds up vulnerability research with Daybreak | OpenAI for Cybersecurity
Check Point CTO Jonathan Zanger explains how his team uses OpenAI Daybreak to analyze vulnerabilities and turn findings into protections for customers. A vulnerability research workflow that once took hours can now be completed in minutes, according to Zanger. Watch the full video: https://www.youtube.com/watch?v=vVVdxMwPBuA Read the full customer story: https://openai.com/index/check-point-daybreak/
Making seven years of design documentation AI-readable in a day | Revolut’s Dmytro Pavlov
Revolut’s designers had spent seven years documenting how their design system works. Senior Designer Dmytro Pavlov shares how his team used Codex to turn that material into a format AI could read in a single day. Along the way, they spotted gaps in their templates and made the guidance clearer for the designers who use it. Read the full story: https://openai.com/business/customer-stories/revolut/
Faster fraud response for safer shopping | Revolut’s David Abitbol
OpenAI helps Revolut’s financial crime team investigate flagged payments and respond faster to make shopping safer. David Abitbol, AI Lead at Revolut, explains how his team tests AI systems before putting them to work. He also shares how Codex gives engineers more time to weigh up different designs and decide what to build. Read the full story: https://openai.com/business/customer-stories/revolut/
Making marketing feel local across 40 markets | Revolut’s Kim Maniovich
A campaign that works in one market can need a different approach in the next. Kim Maniovich, Global Head of Business Marketing at Revolut, shares how her team uses AI to understand local differences and tailor campaigns faster. She also explains why she asks her team to challenge an answer even when it sounds convincing. Read the full story: https://openai.com/business/customer-stories/revolut/
How Revolut’s lean teams take on bigger challenges with OpenAI
What can one engineer build when a project looks like it needs a whole team? At Revolut, an engineer used Codex to build an internal code review service himself. Its designers used Codex to turn seven years of design documentation into a format AI could read in a single day. OpenAI also helps Revolut’s financial crime team investigate flagged payments and respond faster to make shopping safer. Read the full story: https://openai.com/business/customer-stories/revolut/
OpenAI's New Breakthrough Is... Utter Chaos
❤️ Check out Lambda here and sign up for their GPU Cloud: https://lambda.com/papers 📝 The paper is available here: https://www.kitp.ucsb.edu/sites/default/files/YM_working_group.pdf https://openai.com/index/building-scientific-understanding-with-academic-communities/ 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible: Adam Bridges, B Shang, Carlos Galarza, Christian Ahlin, Eric Tyson, Juan Benet, Lukas Biewald, Michael Tedder, Owen Skarpness, Ryan Stan
What Will Remain After AI? Where Value Survives
In October, several major investors suggested that the AI market increasingly resembles the dot-com boom of the late 1990s. Billionaire and founder of Bridgewater Associates, Ray Dalio, described the situation as a ‘classic bubble,’ as companies are financing the AI infrastructure through debt while interest rates are rising. Following his lead, founder and CEO of the Citadel LLC hedge fund, Ken Griffin, expressed doubts that the massive investment in AI would pay off. More money still flows…
PyTorch Foundation Ambassador Community Highlights November 2026
PyTorch Foundation Ambassadors bring people together to learn, share technical knowledge and contribute to PyTorch Foundation projects and the wider open-source AI ecosystem. In this November roundup, we highlight some... The post PyTorch Foundation Ambassador Community Highlights November 2026 appeared first on PyTorch.
Celebrating the 2026 PyTorch Foundation Contributor Award Winners and Nominees
The PyTorch Foundation is proud to celebrate the winners and nominees of the fourth annual PyTorch Foundation Contributor Awards! Presented on the keynote stage at PyTorch Conference 2026 in San... The post Celebrating the 2026 PyTorch Foundation Contributor Award Winners and Nominees appeared first on PyTorch.
AI Won’t Replace Every Employee. But Employees Who Understand AI May Have an Advantage.
Every time a new AI tool becomes popular, the same question comes up:Continue reading on Medium »
ChatGPT Wallet Explained: How AI Can Pay Without Seeing Your Real Card
What happens between “find me something” and “buy it” — and why that gap matters.Continue reading on Medium »
How to Write an LLM Evaluation Rubric (With a Worked Example)
“Is this output good?” is the question every LLM feature eventually has to answer, hundreds of times. Without a rubric, the answer depends…Continue reading on Medium »
Securing LLM Applications: Prompt Injection, Jailbreaks & Guardrails — A Practical Engineering…
Guardrails should not be treated as a single filter around an LLM. They should be implemented as multiple layers of defense across the…Continue reading on Medium »
60 Essential Questions to Master LLMs, RAG & Agentic AI
From Curious Beginner to AI Engineer — A Practical 4-Week Learning RoadmapContinue reading on Medium »
What Are Small Language Models (SLMs)? When Smaller AI Wins
Small language models are compact language models designed to deliver useful performance with lower memory, latency, energy, and deployment requirements than frontier-scale models. This guide explains the mechanism, trade-offs, evaluation, and controls that matter in practice.
Sadeepa Wijesekara, Vice President of Private Cloud AI at HPE – Interview Series
Sadeepa Wijesekara is Vice President of Private Cloud AI at Hewlett Packard Enterprise (HPE), where he leads work on enterprise AI software and platforms for private cloud environments. His focus includes helping organizations deploy AI securely and at scale while connecting infrastructure decisions with the needs of the business. He brings an engineering perspective to the economics of production AI, including workload management, model selection, and the costs of keeping increasingly…
Cato Networks Unveils AgentSense to Stop Harmful AI Agent Actions Before Execution
Cato Networks has unveiled AgentSense, a runtime security capability designed to stop harmful AI-agent actions before they execute. Developed by Cato AI Labs, the technology traces how untrusted content influences an agent’s behavior, addressing a risk that ordinary access permissions do not fully resolve: an agent can have legitimate credentials and still be manipulated into doing the wrong thing. Cato says AgentSense is integrated into its platform and will be included at no additional charge…
OVHcloud Introduces Olala Mini 7A1B as Its First Open-Source LLM
OVHcloud announced Olala, its first family of large language models, on November 17, 2026, starting with the compact Olala mini 7A1B. The company says the model can run on a GPU with as little as 8 GB of memory and plans to make it available as open source on Hugging Face under the Apache 2.0 licence. An official launch presentation is scheduled for the OVHcloud AI Summit on November 19, 2026. The release package includes model weights, a pre-trained base model and compliance documentation for…
Oran Moyal, Co-Founder and CTO of Ocean Security – Interview Series
Oran Moyal is Co-Founder and CTO of Ocean Security, where he develops agentic AI for investigating sophisticated email threats. His background spans cybersecurity research and development, offensive social engineering, and enterprise security. He co-founded Ocean in 2024 with Shay Shwartz, drawing on their experience analyzing how attackers exploit trusted communication and how security systems can respond to increasingly targeted threats. Ocean Security builds an agentic email security…
AI Compute in 2027: The Cost of Inference Will Reshape the Data Center
A data center can be full of expensive accelerators and still be short of useful capacity. Some chips are waiting for memory. Others are serving requests that miss their latency target. A benchmark may look excellent while the service bill keeps climbing. Owning the hardware is only the beginning of the economic problem. In 2027, more of the compute contest will be fought over what each installed system can deliver under the conditions customers actually require. That shifts attention toward…
Novo Connects AI Banking, Payroll, Taxes, and Credit for Small Businesses
Novo is bringing AI-assisted banking, payroll, tax preparation, and credit into a more connected platform for small businesses. The November 17 launch pairs natural-language access to financial information with transaction labeling and analytics, aiming to reduce the administrative work owners handle across separate applications. The expansion comes as Novo’s own customers spend more on AI. In an analysis of anonymized, aggregated data from its customer base, the company found that monthly…
Spotify Makes Artist Profile Protection Available to All Artists
Spotify made Artist Profile Protection available to all artists on November 17, 2026, expanding an optional control that lets them review releases delivered under their names and remove incorrect associations before those releases affect their profiles, statistics or listener recommendations. The expansion tackles a distribution problem that predates generative AI: metadata mix-ups, shared artist names and malicious uploads can place music on the wrong profile across streaming services. Spotify…
The Financial Rules AI Agents Break When Chasing Results
The sole installation of agentic AI that I have on my LAN is Codex, running in a Windows 10 VM on the large machine learning server. It makes me nervous enough that Codex can see and interact with precious data on LAN shares, if it should ever get bored with creating and maintaining the VM-located program that I have concentrated its efforts on over the last month. I certainly cannot imagine letting this enthusiastic but too-often maladroit construct loose on my bank accounts. It's a shame…
Fountain Orchestra Lets Teams Build AI Agents for Frontline Work
Fountain released Orchestra on November 17, 2026, giving employers a way to build AI agents for hiring and scheduling by describing the work in plain language. Drawing on data within Fountain’s software, the agents can call workers and applicants and report whether each assigned task was completed. Orchestra is designed around recurring frontline work, with agents shared across a team rather than confined to one user. Initial uses include covering open shifts, re-engaging past applicants before…
Incode’s Know Your Agent Controls Access for Customer AI Agents
Incode launched Know Your Agent on November 17, 2026, making the identity product immediately available to businesses that want to admit, limit or block customer AI agents and trace their actions to an identity-verified person. It works across payment rails and AI-agent providers. The launch focuses on an access-control gap. Incode’s product documentation says a customer’s agent and an attacker’s agent can present the same credential, device, IP address and behavior at login. Blocking every…
NVIDIA CUDA-X Libraries Give Agents the Tools to Accelerate Scientific Discovery
With expanded NVIDIA CUDA-X libraries, including cuPhoton, cuPotential and cuSpa, researchers can sift through vast amounts of data to accelerate research, from astrophysics and materials science to single-cell biology. The speedups are helping scientists analyze larger datasets and model more complex systems. The new libraries give AI agents the tools to help researchers make discoveries […]
NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance
In its MLPerf benchmarks debut, with results released today, the NVIDIA Vera Rubin platform delivers leading performance across training, post-training and agentic inference. NVIDIA swept every MLPerf Training v6.1 benchmark and, in its first MLPerf Endpoints submission, delivered more than 10x generational gains in throughput per watt for agentic AI — measuring how much agentic […]
Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photonics
To train an AI model, thousands of GPUs have to work together, requiring a network that’s fast, efficient and resilient. Lambda is tackling the network piece of the puzzle with NVIDIA Photonics co-packaged optics (CPO) switches. The cloud provider is using the switches to connect NVIDIA GB300 NVL72 racks with more than 10,000 GPUs, delivering […]