Google DeepMind's Genie 4 world model, now with native audio, a trust framework for agents that improve themselves, and a realistic multi-agent diagnostic environment, lead today. In the news: Anthropic's IPO filing, Meta's plan to raise up to $95 billion, Cerebras and Oracle's planned wafer-scale inference service, and Akamai's new agent execution environment.
Anthropic filed paperwork for a US initial public offering on October 26, disclosing steep losses alongside rapid revenue growth. Reuters reports the filing is being watched as a test of whether huge AI infrastructure spending can turn into durable profit, with a valuation targeted around $1.5 trillion.
Meta plans to raise up to $95 billion in its largest-ever equity fundraising, Reuters reports, to fund the expansion of its AI infrastructure. The plan arrived alongside Anthropic’s IPO filing, in a week when markets are weighing how much capital the AI buildout still requires.
Cerebras and Oracle announced on October 27 that they plan to offer Cerebras wafer-scale AI inference through Oracle Cloud Infrastructure in 2027. The pitch is speed for multi-step agents: fewer data transfers between chips could make each reasoning step faster, though the companies caution that features, timing and pricing can change.
Akamai introduced Akamai Inference and Akamai Agent Sandbox on October 27, adding managed model serving and isolated execution for AI agents to Akamai Cloud. The idea is to give enterprises one place to serve open-weight models, ground them in their own data, and run agent code inside enforced boundaries.
Genie 4 is Google DeepMind’s new world model for real-time interactive simulation, announced on October 27 as a research preview for trusted testers. It builds on the Gemini family, with Gemini Omni anchoring how simulations look and behave, and it generates native audio on the fly, including ambient sounds, music and interactive dialogue. Users can steer a running scene in natural language, changing the weather, moving the camera or introducing new characters, and it is coming soon to Google AI Ultra subscribers in Project Genie and to early access partners in the Gemini API and AI Studio.
World models are moving from short video clips toward places you can inhabit and redirect, and sound is the missing half of that illusion. Genie 4 matters less as a demo than as a test of whether simulated worlds can stay coherent while a user keeps changing the script.
Large-model systems increasingly change themselves through experience and feedback, touching models, prompts, memories, tools and workflows. Recursive self-improvement goes further, revising the mechanisms that govern future updates. This paper maps the foundations, risks and evaluation problems of that idea, warning that reported gains may reflect flawed evaluations, old capabilities may deteriorate, behavioral constraints may weaken, and compromised artifacts may persist into successor systems.
Self-improvement is exciting precisely because it escapes human review, and that is also why it needs a trust framework first. This survey gives the field a shared vocabulary for the failure modes, so claims about an agent improving itself can be checked rather than admired.
Medical agents are usually trained and tested on static question answering, which gives them little practice in actively gathering information, choosing examinations or interacting dynamically with patients. ClinVerse is an interactive, real-data-grounded multi-agent environment built on ClinPatient, a dataset of 150,000 diagnostic cases derived from real electronic health records, with six specialized clinical roles coordinating around a case.
Diagnosis is a process, not a quiz, and benchmarks that only score final answers reward agents that guess well. An environment that scores how an agent investigates could change what medical agents are actually optimized for.
PET scans reveal metabolic information useful for tumor detection and staging, but PET is far less accessible than CT because of cost, infrastructure and injected radioactive tracers. This paper proposes a domain-adapted latent diffusion framework for CT-to-PET translation that transfers natural-image priors using 21,930 paired CT and PET slices, with a signed power transform to handle extreme intensity values.
If a CT scan can stand in for a PET scan, even partially, clinics without PET access gain a metabolic view they could not otherwise afford. The practical question is validation: synthetic PET is only useful if clinicians can trust where it is confident and where it is not.
Diffusion models are strong priors for accelerated MRI, but modeling high-resolution images needs large, costly networks. This framework uses cascaded diffusion models as multiscale priors: a small unconditional network recovers a coarse image from the center of k-space, and a small conditional network fills in fine detail, since detail given a coarse image is local and well conditioned.
Faster MRI means shorter scans, less motion artifact and more patients per machine, and the gains here come from structure rather than scale. Smaller, well-conditioned networks are also easier to audit, which matters when a reconstruction will guide a diagnosis.
Clinical expertise is distributed across institutions, yet medical AI usually aggregates data or parameters rather than experience. This perspective argues for decentralized agentic systems in which distributed AI agents operating across healthcare settings aggregate reasoning experiences instead of datasets, shifting the paradigm from data aggregation to experience aggregation.
Hospitals cannot simply pool patient data, so a credible way to share what agents learn, without moving the data itself, would unlock collaboration that privacy rules currently block. The open question is governance: who checks an experience before it becomes part of another hospital’s agent.
Language agents increasingly store procedural knowledge in reusable skills, and optimizing those skills is a direct route to recursive self-improvement, since the optimizer itself can be encoded as a skill. Existing optimizers follow hand-designed procedures or relearn a fixed meta-skill per task. Skill∞ instead lets a meta agent follow a meta-skill to optimize task skills and optimize the meta-skill itself with the same procedure.
If the way an agent improves can itself improve, progress could compound rather than plateau at the first good procedure. The risk, as the companion survey above notes, is that compounding also compounds errors, so the meta-skill needs the same scrutiny as the skills it produces.
Most fairness benchmarks audit one sensitive attribute at a time, but harms at the intersection of several attributes can be invisible to marginal evaluation. ComFair converts existing fairness data into 479,066 controlled dual- and multi-attribute instances, each paired with its task-equivalent source, so auditors can measure how added attributes shift performance or bias scores across recognition, selection, continuation and conversation tasks.
Fairness failures rarely arrive one attribute at a time in real life, and a benchmark that only checks attributes separately can certify a model that still fails the people at the intersections. A configurable intersectional audit is a practical tool for anyone deploying language models to diverse users.
Group fairness in retrieval-augmented generation is usually measured in aggregate, which can hide topic-level disparities: a system might favor women on human rights questions and men on questions about actors, looking neutral overall while being unfair within each topic. This paper studies fairness within each topic and adapts the reranker separately per topic, using preference pairs built from the generator’s own output.
Aggregate fairness numbers are reassuring and often misleading. Checking fairness topic by topic is a small methodological change with a large practical payoff: it finds the biases that users actually encounter when they ask about a specific subject.
Counterfactual explanations show users what would need to change for a model to decide differently, but generating them well is expensive. CounterFlow uses conditional flow matching with target-conditioned vector fields to steer a source image directly toward a user-specified target class, producing realistic semantic changes with substantially lower computational overhead than GAN- or diffusion-based approaches, and reports state-of-the-art results on standard benchmarks.
Explanations only help if people can act on them, and a counterfactual that shows a realistic, minimal change is far more actionable than a heatmap. Cheaper counterfactuals could make this kind of explanation routine rather than a research demo.
On-policy distillation lets a flow-matching student learn complementary skills from specialized teachers on its own generated trajectories, but different qualities, such as aesthetics and text accuracy, may be best supervised at different stages of denoising. This paper studies two timing decisions, when to query a teacher and when to skip supervision, replacing fixed timestep distributions with adaptive choices that cut unnecessary computation.
Distillation is how small models inherit big-model skills, and its cost is often the bottleneck. Teaching at the right moment, rather than everywhere, is a reminder that training efficiency is partly a scheduling problem, not only a scaling problem.