AI/ML News & Innovations Hub

Scaling Agentic Inference Across Heterogeneous Compute with Zain Asgar - #757

In this episode, Zain Asgar, co-founder and CEO of Gimlet Labs, joins us to discuss the heterogeneous AI inference across diverse hardware. Zain argues that the current industry standard of running all AI workloads on high-end GPUs is unsustainable for agents, which consume significantly more tokens than traditional LLM applications. We explore Gimlet’s approach to heterogeneous inference, which involves disaggregating workloads across a mix of hardware—from H100s to older GPUs and CPUs—to optimize unit economics without sacrificing performance. We dive into their "three-layer cake" architecture: workload disaggregation, a compilation layer that maps models to specific hardware targets, and a novel system that uses LLMs to autonomously rewrite and optimize compute kernels. Finally, we discuss the complexities of networking in heterogeneous environments, the trade-offs between numerical precision and application accuracy, and the future of hardware-aware scheduling. The complete show notes for this episode can be found at https://twimlai.com/go/757.

AI Company Safety Practices Fall Short of Public Commitments and Show Structural Weaknesses, as Top Performers Widen the Gap

But in a win for transparency, five leading companies participated in the scorecard's survey for the first time, providing critical new information to the public.

Vector researchers advance AI frontiers with 80 papers at NeurIPS 2025

Researchers from Vector’s vibrant community are presenting groundbreaking work across the full spectrum of artificial intelligence at this year’s Conference on Neural Information Processing Systems (NeurIPS), taking place December 2-7 […] The post Vector researchers advance AI frontiers with 80 papers at NeurIPS 2025 appeared first on Vector Institute for Artificial Intelligence .

Peers Consulting + Technology partners with global player Andersen Consulting

Brazil-based Peers Consulting + Technology has become a collaborating firm of Andersen Consulting. Founded in 2012, Peers Consulting + Technology (Peers) is one of the fastest growing business and technology consultancies in the Brazilian market, recognized by the Financial Times for three consecutive years.

Stop Saying Boredom is Good for Kids

Chronic boredom causes stress, disengagement, and poor well-being in adults. So why do we glorify it for children?

No, AI hasn’t run out of data

AI models’ relationship with our data is getting more dynamic, contextual and private—and the stakes are high The Claim Earlier this year, Elon Musk claimed that ‘all human data for AI training has been exhausted’. Ilya Sutskever, a co-founder of OpenAI, has said the world has reached ‘peak data’. A recent episode of the BBC’s […] The post No, AI hasn’t run out of data appeared first on OpenMined .

How do you identify and reduce artifacts in Neural Vocoders?

I’m evaluating a few neural vocoders (HiFi-GAN, Vocos, etc.) and I’m seeing the same type of artifact across all of them. I’m not sure exactly what it is, and I’m looking for help identifying it and figuring out how to reduce or remove it. Overall, the models can generate very good audio. In some cases the output is almost indistinguishable from the ground truth. But in other cases, there’s a consistent artifact that’s clearly audible. I’m not an audio expert, so I tried to inspect the signals visually. I plotted both the ground truth (GT) and the predicted audio (PRED) as: waveform (time domain) spectrogram (frequency domain) For the GT audio https://drive.google.com/file/d/1u_p9j29khFrnuqVoHxt8dQ9RgTcjyhpc/view?usp=sharing : The waveform looks “healthy” or “full”. The spectrogram also looks clean and well-structured, and the audio sounds perfect. For the PRED audio https://drive.google.com/file/d/14dsXOZR7JmOwnmNG3kFh6p5Cr7nagrBP/view?usp=sharing : The waveform looks much “thinner” than the GT, like it’s missing energy or detail. In the spectrogram, there are visible “smudges” or “stains” that line up with the artifacts I’m hearing. My intuition is that I might need a better loss function that explicitly penalizes these types of artifacts so the generator learns to avoid them. In the model used to generate the PRED audio, I already implemented a Multi-Resolution STFT loss, which helped to some extent, but the artifacts are still present. This seems to happen with multiple…

SEA-LION Inspires Community to Build AI Solutions That Solve Real-World Challenges in Southeast Asia

AI Singapore (AISG), together with its partners and collaborators, recently hosted the inaugural SEA-LION Summit: Powering Southeast Asia’s AI Future, a key platform demonstrating how SEA-LION is activating the...

What is devops? Bringing dev and ops together to build better software

A portmanteau of “development” and “operations,” devops emerged as a way of bringing together two previously separate groups responsible for the building and deploying of software. In the old world, developers (devs) typically wrote code before throwing it over to the system administrators (operations, or ops) to deploy and integrate that code. But as the industry shifted towards agile development and cloud-native computing , many organizations reoriented around modern, cloud-native practices in the pursuit of faster, better releases. This required a new way to perform these key functions in a more streamlined, efficient, and cohesive way, one where the old frustrations of disconnected dev and ops functions would be eliminated. With two groups working together, developers can rapidly roll out small code enhancements via continuous integration and delivery rather than spending years on “big bang” product releases. Devops was born at cloud-native companies like Facebook, Netflix, Spotify, and Amazon; but it’s become one of the defining technology industry trends of the past decade, primarily because it bridges so many of the changes that have shaped modern software development. As agile development and cloud-native computing have become ubiquitous, devops has enabled the entire industry to speed up its software development cycles. Thus, devops has now thoroughly infiltrated the enterprise, especially in organizations that rely on software to run their business, such as banks,…

Summer Fellowship 2025 Wrap Up - What Did Our Fellows Work On? | GovAI Blog

We're proud to highlight what our 2025 Summer Fellows have been working on.

The overthinking problem in AI

Reasoning models can generate seven to 10 times as many tokens as necessary on simple tasks, creating unsustainable costs at scale. Amazon's vision for metacognitive AI could fundamentally shift how models allocate computational resources.

Ray Serve LLM on Anyscale: APIs for Wide-EP and Disaggregated Serving with vLLM

New study reveals AI’s $100B economic impact across Canada, with Ontario leading the charge

TORONTO, [November 25, 2025] – Ontario is the engine of Canada’s artificial intelligence (AI) economy. AI-related jobs here in the province contributed between $42 billion and $52 billion over the […] The post New study reveals AI’s $100B economic impact across Canada, with Ontario leading the charge appeared first on Vector Institute for Artificial Intelligence .

Alliott Global Alliance expands network with Guatemala-based Palomo Abogados

Alliott Global Alliance has expanded its network in South America with the addition of Palomo Abogados, a Guatemala-based law firm. Founded in 1978, Palomo Abogados provides legal solutions to domestic and multinational corporations, regional businesses, and high-profile clients across Latin America.

Product Evals in Three Simple Steps

Label some data, align LLM-evaluators, and run the eval harness with each change.

Where did it all go wrong? A hierarchical look into multi-agent error attribution

Error attribution in Large Language Model (LLM) multi-agent systems presents a significant challenge in debugging and improving collaborative AI systems. Current approaches to pinpointing agent and step level failures in multi-agent interaction traces—whether using all-at-once evaluation, step-by-step analysis, or binary search—fall short when analyzing complex patterns, struggling with both accuracy and consistency. We present ECHO (Error attribution through Contextual Hierarchy and Objective consensus analysis), a novel algorithm that combines hierarchical context representation, objective analysis-based evaluation, and consensus voting to improve error attribution accuracy. Our approach leverages a positional-based leveling of contextual understanding while maintaining objective evaluation criteria, ultimately reaching conclusions through a consensus mechanism. Experimental results demonstrate that ECHO outperforms existing methods across various multi-agent interaction scenarios, showing particular strength in cases involving subtle reasoning errors and complex interdependencies. Our findings suggest that leveraging these concepts of structured, hierarchical context representation combined with consensus-based objective decision-making, provides a more robust framework for error attribution in multi-agent systems.

Proactive Agents for the Web with Devi Parikh - #756

Today, we're joined by Devi Parikh, co-founder and co-CEO of Yutori, to discuss browser use models and a future where we interact with the web through proactive, autonomous agents. We explore the technical challenges of creating reliable web agents, the advantages of visually-grounded models that operate on screenshots rather than the browser’s more brittle document object model, or DOM, and why this counterintuitive choice has proven far more robust and generalizable for handling complex web interfaces. Devi also shares insights into Yutori’s training pipeline, which has evolved from supervised fine-tuning to include rejection sampling and reinforcement learning. Finally, we discuss how Yutori’s “Scouts” agents orchestrate multiple tools and sub-agents to handle complex queries, the importance of background, "ambient" operation for these systems, and what the path looks like from simple monitoring to full task automation on the web. The complete show notes for this episode can be found at https://twimlai.com/go/756.

LyftLearn Evolution: Rethinking ML Platform Architecture

Written by Yaroslav Yatsiuk At Lyft, machine learning (ML) is the engine behind our most critical business functions — from dispatch and pricing optimization to fraud detection and support automation. Our ML infrastructure serves thousands of production models making hundreds of millions of real-time predictions per day, supported by thousands of daily training jobs that keep ML models fresh and accurate. As our scale grew, we faced a classic engineering challenge: the very complexity that powered our platform was becoming a bottleneck to its future growth. We needed to answer a fundamental question: How could we evolve our platform to accelerate innovation for our users while simplifying its underlying architecture? This post explores how we rethought LyftLearn’s architecture to solve this problem. We’ll walk through our transition from a fully Kubernetes-based system to a hybrid platform, combining the simplicity of managed compute on AWS SageMaker for offline workloads with the flexibility of Kubernetes for online model serving. Afterwards, we’ll share the key technical decisions and trade-offs that made this evolution possible. LyftLearn Overview LyftLearn is Lyft’s end-to-end machine learning platform, managing the complete ML lifecycle from model development to production serving. Built to support hundreds of data scientists and ML engineers, it handles the full spectrum of ML workloads at scale. The platform is composed of three integrated products: Figure 1: LyftLear…

Three Years from GPT-3 to Gemini 3

From chatbots to agents

Is the strong duality of the hard-margin SVM really trivially satisfied all the time?

It is widely known that if you were to calculate the maximizer of the dual SVM program (denote as $\alpha^*$ ), then the primal minimizer of the hard-margin SVM program, \begin{aligned}&{\underset {{w} ,\;b }{\operatorname {minimize} }}&&\|{w} \|_{2}^{2}\\&{\text{s.t.}}&&y_{i}({w} ^{\top }{x} _{i}-b)\geq 1\quad \forall i\in \{1,\dots ,N\}\end{aligned} can be found through the KKT conditions, in which case, $$w^*= \sum_{i = 1}^N \alpha^*_i y_i x_i$$ In particular, those $x_i$ corresponding to non-zero $\alpha^*_i$ are the famous support vectors. However, a technical detail that people seem to ignore is that this only holds if strong duality also holds. I struggled with verifying strong duality, so I searched widely online. Essentially, online reference argue that Slater's condition guarantees strong duality always hold for hard-margin SVM. Slater condition in this case can be written as either: There exists a pair $(w,b)$ that satisfies $y_{i}({w} ^{\top }{x} _{i}-b)> 1$ for all $i = 1, \ldots, N$ , or There exists a pair $(w,b)$ that satisfies $y_{i}({w} ^{\top }{x} _{i}-b)\geq 1$ , for all $i = 1, \ldots, N$ The latter is because the constraint functions are affine. If Slater holds, then strong duality holds and we can make the claim $$w^*= \sum_{i = 1}^N \alpha^* y_i x_i.$$ If it doesn't, then this relationship does not hold. I do not think Slater always holds despite many answers in the contrary online (virtually all of them). I suspect they are making a mistake. I think…

Outstanding students at the Federal Competition for Artificial Intelligence

Diagnosing strokes faster, opening the black box of language models, or determining blood sugar levels painlessly – winning teams provide forward-looking impetus with AI projects

When smart AI gets too smart: Key insights from Vector’s 2025 ML Security & Privacy Workshop

Vector Institute’s 2025 Machine Learning Security & Privacy Workshop revealed critical AI safety breakthroughs and concerning vulnerabilities in current machine learning (ML) security methods. This comprehensive analysis covers the latest […] The post When smart AI gets too smart: Key insights from Vector’s 2025 ML Security & Privacy Workshop appeared first on Vector Institute for Artificial Intelligence .

Altman Solon deepens Brazilian presence with new office in São Paulo

Altman Solon, a global strategy consultancy specialising in telecommunications, media, and technology (TMT), has launched a new office in São Paulo – its second in Latin America. The move in the Brazilian market formalizes the firm’s longstanding work with clients in the country – Altman Solon has been advising local companies and investors for over a decade.

Mixed Methods Scenario Development for Human-Vehicle Interaction Research: A Case Study on Winter Driving

Mixed Methods Scenario Development for Human-Vehicle Interaction Research: A Case Study on Winter Driving robyn.cherinka… Wed, 11/12/2025 - 14:59 Scenarios provide a fundamental link between driving simulators and real-world conditions, shaping the extent to which the findings of a user study can be applied to public roads. However, compared to other aspects of study design, scenario development in human–vehicle interaction research tends to receive less deliberate attention. To encourage more methodical scenario generation, this work introduces a mixed methods approach for extracting representative scenarios from an integration of three real-world data sources: aggregated crash statistics, interviews with experienced drivers, and naturalistic driving data. Through a case study on winter driving, we outline the derivation of a nighttime, two-lane road scenario from these data sources and conduct an initial driving simulator pilot study to assess its realism. We hope that this demonstration of scenario generation from quantitative and qualitative data inspires researchers to consider more rigorous methods for scenario design in future work. Read More Image Oct 8, 2025 Human Interactive Driving 1 Minute Read

Robot Learning from Any Images

Robot Learning from Any Images robyn.cherinka… Wed, 11/12/2025 - 14:54 We introduce RoLA, a framework that transforms any in-the-wild image into an interactive, physics-enabled robotic environment. Unlike previous methods, RoLA operates directly on a single image without requiring additional hardware or digital assets. Our framework democratizes robotic data generation by producing massive visuomotor robotic demonstrations within minutes from a wide range of image sources, including camera captures, robotic datasets, and Internet images. At its core, our approach combines a novel method for single-view physical scene recovery with an efficient visual blending strategy for photorealistic data collection. We demonstrate RoLA's versatility across applications like scalable robotic data generation and augmentation, robot learning from Internet images, and single-image real-to-sim-to-real systems for manipulators and humanoids. Video results are available at this https URL . Read More Image Sep 26, 2025 Robotics 1 Minute Read

From Dashboards to Dialogue: Evaluating a Conversational AI Coach for Performance Driving Skill Development

From Dashboards to Dialogue: Evaluating a Conversational AI Coach for Performance Driving Skill Development robyn.cherinka… Wed, 11/12/2025 - 14:50 Learning in domains involving complex motor skills, such as performance driving, often requires feedback that is timely, personalized, and actionable. Yet many drivers rely on video and telemetry data to review their performance without guidance. We explore how conversational AI can support post-drive reflection by integrating LLM-generated coaching into an interactive review interface. In an exploratory within-subjects simulator study (n=16), participants completed laps under two conditions: one with video and data visualizations alone, and another with the same tools augmented with a conversational interface that provided verbal feedback after each lap. Conversational feedback supported short-term improvements in lap time, average speed, and steering control, and was rated as more useful and satisfying—though it also elicited slightly higher nervousness. These results suggest that conversational AI can make post-drive feedback more interpretable and actionable, particularly for drivers reviewing performance data in high-skill contexts like performance driving. Read More Image Oct 4, 2025 Human Interactive Driving 1 Minute Read

SAFE: Multitask Failure Detection for Vision-Language-Action Models

SAFE: Multitask Failure Detection for Vision-Language-Action Models robyn.cherinka… Wed, 11/12/2025 - 14:45 While vision-language-action models (VLAs) have shown promising robotic behaviors across a diverse set of manipulation tasks, they achieve limited success rates when deployed on novel tasks out of the box. To allow these policies to safely interact with their environments, we need a failure detector that gives a timely alert such that the robot can stop, backtrack, or ask for help. However, existing failure detectors are trained and tested only on one or a few specific tasks, while generalist VLAs require the detector to generalize and detect failures also in unseen tasks and novel environments. In this paper, we introduce the multitask failure detection problem and propose SAFE, a failure detector for generalist robot policies such as VLAs. We analyze the VLA feature space and find that VLAs have sufficient highlevel knowledge about task success and failure, which is generic across different tasks. Based on this insight, we design SAFE to learn from VLA internal features and predict a single scalar indicating the likelihood of task failure. SAFE is trained on both successful and failed rollouts, and is evaluated on unseen tasks. SAFE is compatible with different policy architectures. We test it on OpenVLA, π0, and π0-FAST in both simulated and real-world environments extensively. We compare SAFE with diverse baselines and show that SAFE achieves state-of-the-art failu…

STITCH-OPE: Trajectory Stitching with Guided Diffusion for Off-Policy Evaluation

STITCH-OPE: Trajectory Stitching with Guided Diffusion for Off-Policy Evaluation robyn.cherinka… Wed, 11/12/2025 - 14:40 Off-policy evaluation (OPE) estimates the performance of a target policy using offline data collected from a behavior policy, and is crucial in domains such as robotics or healthcare where direct interaction with the environment is costly or unsafe. Existing OPE methods are ineffective for high-dimensional, long-horizon problems, due to exponential blow-ups in variance from importance weighting or compounding errors from learned dynamics models. To address these challenges, we propose STITCH-OPE, a model-based generative framework that leverages denoising diffusion for long-horizon OPE in high-dimensional state and action spaces. Starting with a diffusion model pre-trained on the behavior data, STITCH-OPE generates synthetic trajectories from the target policy by guiding the denoising process using the score function of the target policy. STITCH-OPE proposes two technical innovations that make it advantageous for OPE: (1) prevents over-regularization by subtracting the score of the behavior policy during guidance, and (2) generates long-horizon trajectories by stitching partial trajectories together end-to-end. We provide a theoretical guarantee that under mild assumptions, these modifications result in an exponential reduction in variance versus long-horizon trajectory diffusion. Experiments on the D4RL and OpenAI Gym benchmarks show substantial improveme…

AI Orchestration for Smart Cities and the Enterprise with Robin Braun and Luke Norris - #755

Today, we're joined by Robin Braun, VP of AI business development for hybrid cloud at HPE, and Luke Norris, co-founder and CEO of Kamiwaza, to discuss how AI systems can be used to automate complex workflows and unlock value from legacy enterprise data. Robin and Luke detail high-impact use cases from HPE and Kamiwaza’s collaboration on an “Agentic Smart City” project for Vail, Colorado, including remediation and automation of website accessibility for 508 compliance, digitization and understanding of deed restrictions, and combining contextual information with camera feeds for fire detection and risk assessment. Additionally, we discuss the role of private cloud infrastructure in overcoming challenges like cost, data privacy, and compliance. Robin and Luke also share their lessons learned, including the importance of fresh data, and the value of a "mud puddle by mud puddle" approach in achieving practical AI wins. The complete show notes for this episode can be found at https://twimlai.com/go/755.

Node.js tutorial: Get started with Node

Node.js is a popular and versatile cross-platform JavaScript runtime environment. Node was the first runtime to allow developers to run JavaScript outside the browser, opening a new world of possibilities in server-side JavaScript . Its ease of use, massive ecosystem and performance characteristics have continued to secure its place as one of the most important technologies of the modern web. Anytime you need to run JavaScript on the server—be it for a systems utility, a REST API, data processing, or anything else—Node is an excellent choice. There are newer runtimes, namely Deno and Bun , but Node remains the standard for server-side JavaScript. Also see: 10 JavaScript concepts you need to succeed with Node . Getting started with Node If you haven’t already experienced Node, this article will introduce you. We’ll step through installing Node and the NPM package manager, spinning up a simple web server, and using the Node cluster module to take advantage of multiple CPU cores. We’ll also look at using the NPM package manager to install additional Node modules and other JavaScript packages. And we’ll dip a toe into using a Node framework, in this case the ubiquitous Express server , to create more feature-rich and flexible Node.js servers. Let’s get started! Installing Node and NPM There are a few ways to install Node , including the installer that Node itself provides , but the recommended way is with a version manager. The most common version manager is NVM . This makes it…

Giving your AI a Job Interview

As AI advice becomes more important, we are going to need to get better at assessing it

Building more accountable multi-modal LLMs through spatially-informed visual reasoning

Recent research has demonstrated that debate mechanisms among Large Language Models (LLMs) show remarkable potential for enhancing reasoning capabilities and promoting responsible text generation. However, it remains an open question whether debate strategies can effectively generalize to Multi-Modal Large Language Models (MLLMs). In this paper, we address this challenge by proposing a location-aware debate framework specifically designed for MLLMs to mitigate hallucination without requiring additional external knowledge. Our approach introduces an asymmetric debate structure across both textual and visual modalities. For textual processing, one MLLM instance generates a comprehensive image description while identifying object locations, while a second instance "zooms in" on specific regions of interest to evaluate and refine the initial descriptions. For visual processing, we introduce a novel hybrid attention module that fuses visual self-attention with cross-modal attention between textual and visual information, effectively highlighting critical content regions. The framework incorporates a judge component that evaluates the complete debate process and selects the most reliable output between the two debating instances. Our experimental results demonstrate that this approach substantially reduces hallucination across diverse MLLMs and evaluation metrics. Moreover, the framework serves as a readily integrable complement to existing hallucination mitigation methods. By emp…

Automated composition of agents: A knapsack approach for agentic component selection

Designing effective agentic systems requires the seamless composition and integration of agents, tools, and models within dynamic and uncertain environments. Most existing methods rely on static, semantic retrieval approaches for tool or agent discovery. However, effective reuse and composition of existing components remain challenging due to incomplete capability descriptions and the limitations of retrieval methods. Component selection suffers because the decisions are not based on capability, cost, and real-time utility. To address these challenges, we introduce a structured, automated framework for agentic system composition that is inspired by the knapsack problem. Our framework enables a composer agent to systematically identify, select, and assemble an optimal set of agentic components by jointly considering performance, budget constraints, and compatibility. By dynamically testing candidate components and modeling their utility in real-time, our approach streamlines the assembly of agentic systems and facilitates scalable reuse of resources. Empirical evaluation with Claude 3.5 Sonnet across five benchmarking datasets shows that our online-knapsack-based composer consistently lies on the Pareto frontier, achieving higher success rates at significantly lower component costs compared to our baselines. In the single-agent setup, the online knapsack composer shows a success rate improvement of up to 31.6% in comparison to the retrieval baselines. In multi-agent systems,…

Fission for Algorithms: The Undermining of Nuclear Regulation in Service of AI

A report examining nuclear “fast-tracking” initiatives on their feasibility and their impact on nuclear safety, security, and safeguards. The post Fission for Algorithms: The Undermining of Nuclear Regulation in Service of AI appeared first on AI Now Institute .

Ray Serve: Advancing Flexibility with Async Inference, Custom Request Routing, and Custom Autoscaling

New schools opening in the UAE for 2026/27: Full list, locations, and fees

What is generative AI? How artificial intelligence creates content

Generative AI is a kind of artificial intelligence that creates new content, including text, images, audio, and video, based on patterns it has learned from existing data. Today’s generative models are typically built on foundation-model architectures such as large-language models (LLMs) and multimodal systems, enabling them to carry on conversations, answer questions, write stories, generate code, and produce images or videos from brief prompts. Generative AI is different from discriminative AI , which draws distinctions between different kinds of input. Where discriminative AI answers questions like “Is this image of a rabbit or a lion?”, generative AI instead responds to prompts such as “Describe to me how a rabbit and lion look different from one another” or “Draw me a picture of a lion and a rabbit sitting next to each other” — and in both cases produces text or imagery that, while grounded in the AI’s training data, isn’t just a copy of something that already existed. [ Read next: Large language models: The foundations of generative AI ] Just a few years ago, generative AI was once a novelty focused on chatbots and artistic image generation. Today, it has become a core enterprise technology, and powers everything from content creation and software development to customer support and analytics workflows. But with that power comes a new set of challenges — from model alignment and hallucination to governance and data-integration hurdles. In this article, we’ll look at ho…

A Guide to Solveit Features

An overview of the features of the Solveit platform, which is designed to make exploration and iterative development easier and faster.

Modifying AI Under the EU AI Act: Lessons from Practice on Classification and Compliance

This is a guest post written by legal compliance professionals Øystein Endal, Andrea Vcric, Sidsel Nag, Nick Malter and Daylan Araz (see section about authors at the end), drawing on their experience from running or consulting businesses integrating AI. For any questions or suggestions, please contact Nick Malter at nick@trail-ml.com. Disclaimer: Please note that the […]

Building Africa’s AI Future Together

Tejumade Afonja is a PhD Researcher at the CISPA Helmholtz Center for Information Security in Germany, where her work explores trustworthy AI, generative models, and synthetic tabular data. She is the 2025 Deep Learning Indaba General Chair and previously chaired the 2023 Indaba Application & Selections Committee. Beyond academia, she is the Director of the […] The post Building Africa’s AI Future Together appeared first on Deep Learning Indaba .

Building an AI Mathematician with Carina Hong - #754

In this episode, Carina Hong, founder and CEO of Axiom, joins us to discuss her work building an "AI Mathematician." Carina explains why this is a pivotal moment for AI in mathematics, citing a convergence of three key areas: the advanced reasoning capabilities of modern LLMs, the rise of formal proof languages like Lean, and breakthroughs in code generation. We explore the core technical challenges, including the massive data gap between general-purpose code and formal math code, and the difficult problem of "autoformalization," or translating natural language proofs into a machine-verifiable format. Carina also shares Axiom's vision for a self-improving system that uses a self-play loop of conjecturing and proving to discover new mathematical knowledge. Finally, we discuss the broader applications of this technology in areas like formal verification for high-stakes software and hardware. The complete show notes for this episode can be found at https://twimlai.com/go/754.

Beyond Standard LLMs

Linear Attention Hybrids, Text Diffusion, Code World Models, and Small Recursive Transformers

Announcing Anyscale Runtime for Faster, Cheaper and More Resilient AI, Powered by Ray

Ray Train V2: Unified Distributed Training on Ray

Ray Data: Scalable Data Processing for AI workloads

Anyscale Announces Lineage Tracking for Faster End-to-End Tracing and Debugging of Ray Workloads

Ray Summit 2025: Anyscale Platform Updates

Announcing Ray Direct Transport: RDMA Support in Ray Core (Part 1)

[Paper Analysis] The Free Transformer (and some Variational Autoencoder stuff)

https://arxiv.org/abs/2510.17558 Abstract: We propose an extension of the decoder Transformer that conditions its generative process on random latent variables which are learned without supervision thanks to a variational procedure. Experimental evaluations show that allowing such a conditioning translates into substantial improvements on downstream tasks. Author: François Fleuret Links: Homepage: https://ykilcher.com Merch: https://ykilcher.com/merch YouTube: https://www.youtube.com/c/yannickilcher Twitter: https://twitter.com/ykilcher Discord: https://ykilcher.com/discord LinkedIn: https://www.linkedin.com/in/ykilcher If you want to support me, the best thing to do is to share out the content :) If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this): SubscribeStar: https://www.subscribestar.com/yannickilcher Patreon: https://www.patreon.com/yannickilcher Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2 Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n

RL without TD learning

In this post, I’ll introduce a reinforcement learning (RL) algorithm based on an “alternative” paradigm: divide and conquer . Unlike traditional methods, this algorithm is not based on temporal difference (TD) learning (which has scalability challenges ), and scales well to long-horizon tasks. We can do Reinforcement Learning (RL) based on divide and conquer, instead of temporal difference (TD) learning. Problem setting: off-policy RL Our problem setting is off-policy RL . Let’s briefly review what this means. There are two classes of algorithms in RL: on-policy RL and off-policy RL. On-policy RL means we can only use fresh data collected by the current policy. In other words, we have to throw away old data each time we update the policy. Algorithms like PPO and GRPO (and policy gradient methods in general) belong to this category. Off-policy RL means we don’t have this restriction: we can use any kind of data, including old experience, human demonstrations, Internet data, and so on. So off-policy RL is more general and flexible than on-policy RL (and of course harder!). Q-learning is the most well-known off-policy RL algorithm. In domains where data collection is expensive ( e.g. , robotics , dialogue systems, healthcare, etc.), we often have no choice but to use off-policy RL. That’s why it’s such an important problem. As of 2025, I think we have reasonably good recipes for scaling up on-policy RL ( e.g. , PPO, GRPO, and their variants). However, we still haven’t found a “…