2026-09-26

Anthropic has committed $11.6 billion over seven years to Akamai's cloud infrastructure, a bet on CPUs that could grow to about $20 billion, and in an unusual arrangement, Akamai is giving Anthropic a potential stake of up to 5% of its stock that grows as Anthropic spends more.

Frontier AI models are finishing Alan Turing's World War II codebreaking work.

The tech giant, which allows companies to send large datasets over the internet, said it received a "credible threat" from law enforcement about an imminent attack.

Every time I’ve seen microservices pitched, it sounds great on paper. Independent teams, clean ownership, scale only what you need. Then a year later you’ve got dozens of services, three different deployment patterns, tracing everywhere, and nobody really understands the whole thing anymore. Maybe I’ve just seen bad implementations, but I’m starting to think way fewer companies actually need microservices than we pretend.

The rapid capability gains of frontier language models are widely attributed to improved reasoning abilities, yet this cannot be verified as raw CoT traces in closed-source systems are hidden. By registering a simple custom tool through a standard API feature, we induce frontier models to externalize intermediate reasoning. Because these traces may reflect post-hoc rationalization rather than genuine reasoning, we first evaluate against native CoT on open-source models and extend to closed-sourc

Sparse AutoEncoders (SAEs) offer a promising way to inspect language model representations, but it is still unclear what kind of linguistic structure their latents expose. We use part-of-speech (PoS) categories as a controlled test case to study whether morpho-syntactic information is encoded by individual latents or by structured groups of features. We find that PoS distinctions are highly recoverable from SAE activations, but do not align with one-to-one latent / category mappings. This recove

TL;DR - Swift Flash is a killer model that massively reduces excess reasoning. Try it out! If you haven't seen from my previous comparison posts , I'm a huge fan of the Swift Qwen3.8 models. I've been using 27B since it dropped, and I'm really impressed with the performance and quality (v1.5 is even better). The reduction in overthinking is a huge win, and quality seems to be essentially equivalent in real-world use and benchmarking. The time savings are massive. When UkisAI told me they were pl

Language-model agents increasingly face long-horizon tasks with evolving state, interdependent decisions, and delayed outcomes. Scaling their training requires diverse agentic environments, dependable outcome signals, and low extension cost. Existing generation pipelines commonly construct an environment before defining its outcome rule or annotating its trajectories, leaving dynamics and evaluation to be aligned post hoc. VHD-Play reverses this dependency by sampling and solving a mathematical

Can one language model hand its live memory to another without the receiver rereading the context? We demonstrate useful persistent hybrid-state transfer across one architecture-matched Qwen3.5 4B-to-9B sibling pair. To our knowledge, this is the first demonstrated cross-model handoff of persistent recurrent inference state between differently sized hybrid language models without target prefix replay. Translated attention KV alone leaves a large gap; adding the Gated DeltaNet (GDN) persistent-st

Vendors rolling out AI have an "incentive to deny as many claims as possible."

The funding, which comes from Third Point, Nvidia, and others, will fuel the company's massive AI data center buildout.

Video generation begins in text space by authoring a cinematic screenplay, then materializes into pixels. As contemporary video generators scale to 30 seconds and faithfully follow complex conditions, the textual prompt largely directs the production, planning how actions, camera trajectories, lighting, and sound unfold across multi-shot sequences. In this paper, we present WanPE, a 397B-parameter prompt enhancement model trained on 1.05M real-world videos to master director-level cinematic plan

arXiv:2609.28475v1 Announce Type: new Abstract: Forecasting agents increasingly combine language-model reasoning, retrieval, ensembling, and calibration, but it remains unclear when each behavior should be trusted. We study this question on ForecastBench-style binary forecasting tasks, treating the choice to retrieve, reason, defer to a market prior, or use a historical analog as an observable agent behavior rather than a hidden implementation detail. Our central finding is that mechanism choice

Humans can effortlessly localize the direction of a sound source and integrate it with visual cues for reasoning, yet this remains challenging for embodied agents. In particular, it is still unclear how to effectively evaluate and model spatial audio understanding in embodied settings. To address this gap, we introduce OmniEchoBench, a unified benchmark for spatial audio-visual perception and audio-vision-language navigation. OmniEchoBench comprises six tasks over 197 real-world spatial audio-vi

arXiv:2609.28570v1 Announce Type: new Abstract: Reinforcement learning (RL) is widely used to sharpen reasoning in multimodal large language models (MLLMs), yet its effect on hallucination is uneven. We trace this to two weak points in the \emph{correction chain} from reward to parameter update. At the rollout level, hard queries---those with high semantic entropy---frequently produce unanimously wrong sample groups, collapsing the group-relative advantage to zero exactly where hallucination ris

arXiv:2609.28609v1 Announce Type: new Abstract: Role-playing agents based on large language models have been widely applied in areas such as personalized assistance and social simulation. Recent RL methods typically train on a fixed scenario pool collected before learning begins. This creates a distributional bottleneck: as the agent improves, the scenarios where it performs poorly also change, while the training distribution remains static. Therefore, we propose AdvRole, an adversarial context

arXiv:2609.28554v1 Announce Type: new Abstract: We introduce the Pistis model family, comprising 27B- and 9B-parameter multimodal large language models built on Qwen3.6 and Qwen3.5, respectively, and developed through a general and scalable post-training framework. The framework first establishes a strong foundation through large-scale multimodal supervised fine-tuning (SFT). Building on this SFT foundation, we propose Interleaved Distillation and Reinforcement Learning (IDRL), a novel post-trai

GWM Worlds 2 uses persistent context and timed actions to steer a world model generating video and audio in real time.

Build charts and dashboards from Cursor and Claude Discussion | Link

I've been playing around with various models on the M5 Ultra 256GB 80-core Mac Studio. These are the results over many rounds of agentic inferencing. I'm happy with the performance. Glad to have the large amount of RAM. But it does feel like the GPU is underpowered for this amount of RAM. I'm wondering if a 512GB unit for AI inference makes sense at all - because the GPU will be the clear bottleneck.

AI research agents need reliable knowledge of how their experiments change outcomes. We introduce WhatWorkedBench to measure experimental understanding, the accuracy of predictions about component changes after budgeted experimentation. Agents inspect code, select measurements, and submit a response surface, a table predicting scores for every configuration of component settings. Exhaustive CPU execution supplies reference effects for changing each component while holding the others fixed. These

Just saw this pop up. This might be a fun one for the folks in here!

Every rent-vs-buy thread I read has confident people on both sides, but not many actually show the numbers. So I finally ran the numbers for our own decision. Posting the working here in case it is useful, or feel free to point it out in case someone thinks it's wrong. An 8-GPU HGX H200 server lands somewhere near $320k-$420k, with roughly $370k being a reasonable midpoint. On the rental side, the median on demand H200 price across 34 providers was about $4.40/GPU-hour as of September 18. The $2

We hosted Kev 4B (Jared Palmer's Apache-2.0 fine-tune of Qwen3.5-4B) and ran it side by side with Jev on the same endpoint to see how it compares. We built a fresh set of 362 items published after both models shipped (new arXiv papers, Stack Exchange questions, GitHub issues), with answers taken from the source. A few findings: - Accuracy lands within 2 points on every task, inside the noise at this sample size - Jev is better calibrated and pulls ahead on paraphrase detection (PAWS 87.0% vs 74.

Boom Supersonic CEO Blake Scholl said its new stationary power plants were no longer in Crusoe's near-term plans.

The AI lab had argued multiple violations of its rights, but a divided panel of judges sided with the Trump administration.

Microsoft is giving its Copilot app another facelift.

arXiv:2609.28547v1 Announce Type: new Abstract: Policy interventions propagate through public communication, institutional decisions, and stakeholder responses, yet datasets for financial multi-agent simulation rarely connect these processes to temporally aligned historical evidence. We introduce PAWS, a Policy-driven Agentic World Simulation dataset covering 36 verified U.S. financial and economic policy episodes, 12,727 policy-linked news records, and 65,291 source-grounded stakeholder actions

arXiv:2609.28771v1 Announce Type: new Abstract: Large language models (LLMs) have demonstrated strong capabilities in financial analysis and reasoning, inspiring recent advances in agent-based trading frameworks. While these systems show promise, prior approaches either emphasize long-horizon forecasting or operate as stateless analyzers, limiting their applicability to the demands of trading in complicated settings. To address these gaps, we introduce META (Memory Enhanced Trading Agent), the f

arXiv:2609.28557v1 Announce Type: new Abstract: DNA sequencing pipelines, spanning quality control, alignment, variant calling, and annotation, are now reliably executed by workflow management systems that orchestrate established bioinformatics tools at scale. What remains manual is the decision layer surrounding that execution: selecting quality thresholds appropriate to a sample and platform, adjudicating borderline variant calls, diagnosing anomalies, and determining which findings warrant ex

arXiv:2609.28690v1 Announce Type: new Abstract: Faithful user simulation is fundamental to building, evaluating, and improving interactive AI at scale. However, plausible individual responses do not ensure that simulated users reproduce the intent evolution and outcomes observed in real interactions. We propose TRACER, a multi-turn user simulator that explicitly models users' evolving intent and learns to align simulated behavior with real interaction trajectories. TRACER is trained in two stage

arXiv:2609.28575v1 Announce Type: new Abstract: Long-conversation memory benchmarks increasingly test recall and prompted knowledge updates, and recent work studies evolving user beliefs and memory state. TWIST is a proposed benchmark suite for a complementary, unmeasured property: intervention quality -- whether a deployed memory system, exercised through its own ingest/recall/vet surface, acts correctly at belief change points. Four tracks cover unprompted tension detection, vetting outgoing d

Shows an Incorrect API Key error...which is weird. Nothing on the status page. (EDIT: Added to incident page: ) Per Twitter/X this appears to be widespread

Irrational Analysis:"HBM is a mistake" Former Intel CEO: "HBM is lousy" SK Hynix VP:"HBM is not the final answer to the memory wall problem" "If the stacks get high enough...each core die operates slower than plain old commodity memory" Hot chips 2026 Q&A, Irrational Analysis asks: "You talk about going to 20 levels thick at HBM4 so maybe you are at 4 terabytes per square cm in a stack of 20, so you are talking about having 20% of the bandwidth of one chip [for each] layer of 20. You've diluted

Omni-modal large language models (OLLMs) jointly process vision, audio, and text, yet their modality bias under cross-modal conflict remains underexplored. Existing benchmarks conflate two distinct forms of evidence within a single modality: perceptual signals (e.g., a photograph or recording of a dog) and propositional signals (e.g., the declarative claim "this is a dog"), such that any measured modality bias is inherently confounded with evidence-form bias, precluding clean attribution to eith

Something about the Oracle numbers has been bothering me and I think I finally put my finger on it. 21,000 cuts this year. $1.8 billion severance bill. Another 800 scheduled for November 13 according to WARN filings. All happening alongside enormous capex commitments for AI data center buildout. The public framing is AI-driven restructuring. But if you actually look at the cash flow, the cuts aren't a consequence of automation replacing those roles. They're how the capex gets funded. You cut ope