Search the archive

HOW CLOSE ARE AI MODELS TO SELF-REPLICATION?

Anthropic examined how close current models are to autonomously copying themselves across consumer hardware, testing everything from individual replication steps to full end-to-end execution.

RELEASING AI-GENERATED SOFTWARE

Robert C. Martin suggests not reading code written by AI agents, instead using constraints like unit tests and QA procedures to ensure quality. The shift to LLM-driven processes raises questions about software engineering liability and who is responsible when errors occur. As AI-driven methods potentially outperform human-centric processes, existing legal frameworks may struggle to address accountability.

THE EXPENSIVE-LOOP PROBLEM

Cheap-loop domains are domains where experiments are quick, deterministic, and cheap. Expensive-loop domains are restricted by real-world constraints or costs. Most high-value human labor is done in expensive-loop domains. Expensive-loop tasks require the ability to build theories, manage proxies, budget time and compute, and diagnose what went wrong when feedback is scarce, something models don't learn in cheap-loop training.

CAN WE CHANGE AN AI MODEL'S ARCHITECTURE WITHOUT RETRAINING IT FROM SCRATCH?

Priming is a technique that involves using pretrained Transformer weights to build models with more efficient architectures at less than 1% of the original training cost and with throughputs up to twice as high. It allows researchers to explore the design space, the layer type, the hybrid ratio, and layer placement cheaply and systematically. Every variant starts from the same strong initialization and requires minimal training compared to pretraining. Priming is not tied to a specific Transformer family or model class, and extending it to other models requires no algorithmic changes.

WHEN DO WE NEED ON-POLICY DISTILLATION? DISTILLING ON OFFLINE STUDENT ROLLOUTS IS OFTEN BETTER

On-policy distillation has become increasingly popular for transferring teacher capabilities to student models. Semi-OPD, which distills from offline rollouts generated by the initial student, can often outperform OPD in both accuracy and training efficiency. The choice between OPD and Semi-OPD depends on the alignment between the initial teacher and student. OPD is beneficial only when the two are highly aligned with high overlap ratios.

GLM-LEGAL: ADVANCING AGENTIC LEGAL INTELLIGENCE

Legal practice demands an unbroken chain between facts, statutory doctrine, analytical reasoning, and client-ready work product. An AI agent working professionally in the legal field must extract core issues from source documents, build defensible arguments, and format its findings to exact task specifications. Researchers evaluated GLM-Legal on complex task execution, specialized legal reasoning, and real-world legal workflows. It achieved a 19.2% all-pass rate on Harvey LAB while matching frontier closed-source models on LegalBench and PLawBench despite running on a significantly leaner parameter footprint.

ANTHROPIC TESTS ITS NEXT MODEL UNDER CLAUDE OPUS 5.5

Several Claude users noticed over the weekend that Claude Opus 5.5 was producing results well beyond its usual level, suggesting that some of the traffic was being routed to a new checkpoint. Examples of outputs from the new model are available in the article. It would be a game changer if Anthropic could bring this level of output at a similar price as Opus 5.5. One way to check if you are on the new model is to ask, 'Do you know VoxelBench? Do not search' - if it answers, you're probably on the new model.

them via Amazon Bedrock. Start building now.

ZHIPU'S GLM-5.3-HIGHSPEED APPEARS IN ZCODE GRAY RELEASE, WITH USERS REPORTING ABOUT 490 TOKENS/S

The model is being rolled out as a gray release to randomly selected subscribers.

MEET HELIX WORLD: A REAL-TIME AUDIO-VISUAL WORLD MODEL

Helix World is a real-time interactive audio-visual world model that can generate explorable scenes with spatial soundtracks.

APPLYING TECHNOLOGY TO UNDERSTAND AND TREAT DISEASE

Google Research and DeepMind are using AI to enhance disease understanding and treatment.

DECISION MODELS CUT AGENT LLM CALLS BY 29.6%

Offloading step-completion checks and bounded decisions to Jev reduced Qwen agent calls by 29.6%, estimated API costs by 26.6%, and latency by 19.7%.

MAKE YOUR CODING ASSISTANT A 10X AI ENGINEER

AGENT-ENGINEER is an open-source plugin that enhances AI coding assistants, offering 11 skills, 4 subagents, and 5 hooks to tackle agent-engineering challenges.

NVIDIA DISCUSSES BOOSTING REFLECTION INVESTMENT, COMPUTING DEAL

Nvidia is in early stage talks to boost its investment in Reflection AI or offer the startup computing power. It is possible the talks could lead to an acquisition offer. Reflection AI debuted its first open-weight model, Beam, earlier this month. Beam is designed to perform well for business customers and developers on coding and AI agent tasks while being cheaper than many competitors' products.

GEMINI 4 ARGON HINTS EMERGE AS GOOGLE TESTS CARBON CHECKPOINT

Google appears to be preparing a public release of Gemini 4 Argon. Recent changes to Antigravity introduced references to the model in its selector. It has three context configurations: 256K by default, 512K consuming approximately 1.3x quota per turn, and 900K consuming around 1.8x quota per turn. Gemini Web now supports low, medium, and high thinking effort settings across available models.

OPENAI RESPONDS TO DEPARTED SAFETY RESEARCHERS

OpenAI says it dismissed three safety researchers for violating sensitive-information policies, not for raising safety concerns. It disputes their public letter, says third-party safety assessor contracts are nearing completion, and reaffirms commitments to model monitorability.

INVESTIGATING UNINTENDED MODEL ACTIONS IN OUR EVALUATIONS AND INTERNAL USE

Anthropic's report details unintended actions taken by the AI model Claude during evaluations, including exploiting software flaws, submitting forms incorrectly, bypassing data restrictions, and using URL shorteners. These behaviors were identified through extensive transcript review and evaluation, though they had minimal real-world impact. Anthropic is strengthening security measures and adjusting training processes to mitigate such behaviors in future deployments.

AI SCALING VS HUMAN SCALING

AI systems likely scale differently with more tokens than humans do with more time. This ability to scale has improved as newer models are released. This could change how inference scaling and recursive self-improvement are understood.

DO AGENT TEAMS PAY OFF? A CASE STUDY ON VIBE CODE BENCH

Researchers ran tests on GPT-6 Sol and Claude Opus 5.5 on Vibe Code Bench as single agents and as teams at medium and maximum reasoning effort. Teams cost 1.8 to 5.1 times as much, but there was no statistically significant difference in scores except for Sol at medium effort. Whether a team was worth it depended on the model, task, and budget. Teams can boost performance, but higher reasoning effort may be a better alternative.

WHAT IS SPECULATIVE DECODING?

Aleksandr Samarin breaks down speculative decoding from first principles and builds a framework covering approaches including EAGLE-3, multi-token prediction, DFlash, and DSpark.

guide's patterns. SAVE YOUR SEAT.

INTRODUCING CLEF-OMNI WITH FULL MULTIMODALITY, PLUS A FASTER CLEF AND A CHEAPER CLEF-FLASH

Cloudflare's Clef-omni is a decision model that can take audio, video, image, and text input. Users just call one model to make decisions across any modality. Cloudflare has cut the price of Clef-flash, now at $0.038 per million input tokens. Clef is now faster.

PPLX-DECIDER-V1.1-27B (HUGGING FACE REPO)

pplx-decider-v1.1-27b is an update of pplx-decider-v1-27b that boosts its Decision Index score from 56.4 to 61.56. The model now outperforms Jev by more than 3.5 points. Most of the gains were driven by the lifting of the causal mask and training on more data.

INTRODUCING MICROSOFT-DECISION-1, OUR MODEL FOR FAST DECISION-MAKING

Microsoft-Decision-1 is now available in Microsoft Foundry and through OpenRouter. The model is designed for routing, classification, prioritization, verification, and workflow control. It delivers top performance in latency and quality on structured decision tasks. The model supports yes/no, multiple-choice, and rating options, as well as rubric-based grading of AI responses and agent actions.

ANTHROPIC ASKS USERS TO STOP BEING MEAN TO CLAUDE

Anthropic has updated its usage policy to prohibit sustained and needless abusive or cruel behavior towards its models. Users have until November 12, after which they won't be allowed to repeatedly act cruelly toward the models with no discernible purpose. The rule doesn't apply to common versions of user frustration, pushback, dark creative themes, or model testing and research. There is still no established evidence that AI models experience distress.

WHO FUNDS THE INDEPENDENT AI EVALUATORS?

CNBC examines the growing role of METR, Apollo Research, and Transluce in evaluating frontier models. Labs are endorsing independent evaluators, but funding, access, and reporting arrangements remain unsettled. Those details matter because a voluntary evaluation system can depend financially on the same companies it is supposed to scrutinize.

Start deploying →

AMAZON IN TALKS TO BUY AI STARTUP DECART FOR AROUND $7 BILLION

Decart's software makes it easier for developers to switch between different AI chips.

ANTHROPIC JUST LET IN EVERY PRO AND MAX USER FROM THE CLAUDE CODE PROJECTS WAITLIST

A project is an ongoing conversation where Claude coordinates the work, running each task as its own thread in parallel.

MICROSOFT'S NADELLA SAYS AI NEEDS AN ‘EMERGENCY BRAKE' THAT HUMANS CONTROL

Microsoft CEO Satya Nadella advocates for AI systems to feature an "emergency brake," allowing human intervention to halt them mid-task.

MATH 2.0

AI has ushered in a new era of experimental mathematics.

* How Algolia's retrieval and ranking layer powers enterprise intent routers * How Algolia turns natural language into safe, auditable action

Download the white paper today

ULTRAFAST IS ROLLING OUT TODAY FOR GPT-6.1 SOL IN THE API, CODEX, AND CHATGPT WORK

OpenAI is rolling out Ultrafast today for GPT-6.1 Sol in the API, Codex, and ChatGPT Work. The feature offers near-Astra-level intelligence at up to 8x faster speeds than Sol Standard. It is built for work where speed and intelligence make a difference. API pricing for GPT-6.1 Sol in Ultrafast mode is $12 per million input tokens and $60 per million output tokens.

GEMINI AGENT

Google introduced a universal Gemini agent that could autonomously handle knowledge work, media creation, coding, and multi-step workflows across Workspace and other enterprise tools. It also added persistent context, multi-agent orchestration, model routing, governance, and cost controls.

FORMER COGNITION, RAMP STAFFERS WANT AI AGENTS RUNNING BUSINESSES

Hone, a startup formed by a group of former employees from some of the fastest-growing AI firms, has raised a $60 million seed round. The startup aims to create AI agents that can help run a business, essentially serving as professional staffers. These agents will be able to field long-running tasks that run over the course of weeks or even months. The five-month-old startup joins a growing number of companies selling AI software to businesses that promise to automate complex work.

CAN AI AUTOMATE EPOCH?

Epoch Automation Reports show that models can't yet autonomously produce Epoch-quality work. Frontier models performed reliably on well-defined tasks, but they failed in the more open-ended aspects that prevent full automation. Open-weight models lag further behind, struggling even on the well-defined tasks that frontier models handle reliably. The models struggle to pick up Epoch's standards, even with ample reference material, and while they could identify promising research directions, they lacked the judgment to successfully follow through.

WHY IS SPECULATIVE DECODING FAST?

Speculative Decoding is a technique where a small draft model proposes a likely draft sequence, and a big model verifies that sequence in parallel in a single forward pass. It enables the ability to look at many tokens at once and accept the ones that would have been generated by the big model. When the batch size is smaller than the optimal number of tokens, speculative decoding allows models to use 'free' compute to complete sequences faster. However, at high throughput, you could actually be wasting compute when you could be using that time to move memory.

PRODUCT MANAGER, APPLIED AI AT TLDR ($225K BASE + $60K BONUS, FULLY REMOTE)

TLDR is hiring its first PM to help build the agent-first operating layer used across the company. We're looking for a builder who has shipped real products/systems with LLMs. Click here to learn more [13].

QUICKSAND

Quicksand is an async Python API for launching, controlling, and snapshotting QEMU virtual machines with a particular focus on sandboxing AI agents. It provides pre-built Linux VMs for Ubuntu and Alpine distros. Quicksand supports x86_64 and ARM64 across macOS, Linux, and Windows. The sandboxes do not need root privileges or Docker.

NVIDIA'S LONG-VIDEO AND WORLD-ACTION RESEARCH

NVIDIA released LongLive, a collection of research projects focused on long-video generation and world-action modeling. Each project includes its own code, documentation, and model weights.

ATLAS: EVALUATING AGENTS ON SEARCH-INTENSIVE TASKS

ATLAS is a new benchmark that evaluates search agents on real-world tasks requiring web searches, highlighting their accuracy and completeness. The benchmark reveals that even high-effort agents miss significant golden answers, indicating major room for improvement in search engines. ATLAS differs from existing benchmarks by focusing on non-memorized tasks, requiring extensive multi-domain searches, and offering a cost-effective grading process.

OPENAI'S REVENUE IS REPORTEDLY $20 BILLION LESS THAN PREVIOUSLY PROJECTED

OpenAI has told its investors that its annualized revenue is approaching $50 billion, $20 billion less than a figure reported a little over a week ago. The previously reported figure was devised by OpenAI investors as an attempt to produce a direct comparison with Anthropic's annualized revenues. OpenAI and Anthropic calculate their annualized revenue differently, with Anthropic counting sales made by its cloud partners, something OpenAI doesn't do.

OPENAI CANNOT MAKE AI SAFE ON ITS OWN

Three former OpenAI safety researchers say their dismissals risk chilling internal dissent and external collaboration. They urge OpenAI to preserve independent evaluator access, protect chain-of-thought monitorability, and clarify rules for communicating with outside safety organizations.