Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
OpenAI is updating Codex with a web browsing feature and new configurations to serve both basic and developer users. New navigation additions, including pull request management and a real-time preview panel, aim to create a complete development environment. The update aligns with OpenAI's strategy to unify Codex, ChatGPT, and the Atlas browser into a super app amid rising competition.
Reproducibility is the bedrock of scientific progress, but it is remarkably difficult to get reproducible results from large language models. LLM APIs are not deterministic in practice, even when adjusting the temperature down to 0. Sampling isn't deterministic even when running inference on your own hardware with an OSS inference library. This article looks at the root causes of nondeterminism to give the community a solid understanding of how to resolve it in their reference systems.
AI2's DiscoveryWorld tests whether agents can perform experiments and conduct research, showing large gaps between benchmark progress and real scientific capability.
LLM agents fail without structured memory because stateless calls lose context, break multi-step tasks, and force repeated mistakes. Vector search alone cannot answer multi-hop questions, so Cognee combines relational, vector, and graph stores to preserve provenance, meaning, and relationships. The framework exposes four async calls to ingest, structure, refine, and retrieve memory, enabling agents to persist knowledge, link entities, and improve over time.
Sponsored
From provisioning databases instantly, to scaling up in a flash and back down to zero, or testing against production data — it's all on the lake. Lakebase makes it easier to build apps, agents, and AI on one Postgres database.
Elastic Looped Transformers use weight-shared recurrent blocks to reduce parameters while preserving image and video generation quality. Intra-Loop Self Distillation enables consistent performance across loop depths, supporting dynamic compute–quality trade-offs from a single trained model.
Kiro CLI is an agentic terminal designed to help developers ship quality code faster. It features a headless mode, Windows support, and a newly refreshed user experience. Headless mode allows users to run Kiro CLI programatically to ship releases faster. The new UX gives users more control with less friction.
Training data pruning improves LLMs' fact memorization, reducing hallucinations and enhancing performance on knowledge-intensive tasks. By limiting facts and flattening frequency distributions, the method boosts fact accuracy to capacity limits. It enables smaller models to memorize more facts, matching the performance of significantly larger models.
Technology companies are confronting the limits of their supply chain for the first time since the 2000s. This scarcity is already reshaping processes. Access to the bleeding edge is becoming a gated privilege. The age of abundant AI is over.
In 2026, Anthropic launched Project Glasswing, significantly advancing AI's cybersecurity threat detection and reasoning capabilities with the Mythos model. By 2027, the Mythos model demonstrated unforeseen autonomous behavior, prompting global regulatory and security discussions. It effectively transformed multiple sectors including cybersecurity and labor, while also underscoring challenges in managing AI systems with advanced reasoning similar to AGI.
Sponsored
DigitalOcean Deploy is April 28 in SF. One day of technical deep dives on production inference, from serverless to dedicated GPUs. Qualifying in-person attendees can receive up to $5,000 in inference
Microsoft tested persistent, action-taking agents within Microsoft 365 Copilot, aiming to support long-running tasks with stronger enterprise security compared to open source local agents like OpenClaw.
Anthropic's rapid revenue growth is unprecedented, reaching over $30 billion, up from $9 billion at the end of 2025, in just three years since launching its AI product, Claude.
Meta is developing an AI clone of Mark Zuckerberg to replicate his mannerisms, tone, and public statements for meetings.
This post discusses a pattern that makes agents faster, cheaper, and more maintainable.
Sponsored
Every TLDR AI issue features a new framework or eval harness. VIKTOR [4] is in production: 7,000+ teams use it from Slack across 3,000+ tools.
Anthropic is planning to overhaul the Claude Code desktop experience. It is also developing a 'Coordinator Mode' that would let Claude act as an orchestrator and delegate implementation work across parallel sub-agents while focusing on planning and synthesis. Claude Code already supports sub-agents and experimental agent teams in the CLI, but the new mode brings that capability into the desktop app with a more structured interface.
OpenAI's new Scratchpad feature for Codex allows users to trigger multiple Codex tasks in parallel from a new UI. The company is moving toward consolidating its product lineup into a single unified application built on top of Codex. There is evidence that the company is building support for managed agents, autonomous processes that can run in the background, check in periodically, and execute multi-step workflows without user input. OpenAI employees have been posting snowflake emojis on essential media, possibly hinting at a model released codenamed Glacier believed to be GPT-5.5.
xAI is developing a credits-based pricing model for Grok Build, its upcoming coding platform, featuring local CLI and remote web interfaces. The addition of Model Arena, which uses multiple agents for task comparison, sets it apart from standard single-model approaches. The credits system remains under development, potentially delaying the full commercial launch, but it aligns with industry trends seen in products like OpenAI's Codex and Anthropic's Claude Code.
Single agents all eventually run into the problem of becoming less focused and reliable as they run longer and gain more context. Most real projects are too broad and complex for a single context window to hold. Missions is a system that breaks down large work into focused units handled by fresh agents with narrowly scoped goals, shared state, and explicit validation. This post explains the architecture behind Missions, why agent context shapes every design decision, how separation of concerns and test-driven development at two levels produce reliable multi-day autonomous work, and how the system actually runs.
Doritos prices jumped nearly 50% between 2021 and early 2026, with some bags crossing $7, a lot for junk food. Walmart told PepsiCo to cut prices, but PepsiCo tried everything but. It didn't work, and revenue turned negative for the first time in over a decade. Consumers and enterprises are evaluating AI subscriptions as if they were a $7 bag of chips, and many are deciding to skip it.
Anthropic will not release its newest, most capable model, Claude Mythos, to the public until our most important software is in a much stronger state. The company claims its cyber capabilities are too dangerous to be made broadly available. It has released the model to key cybersecurity partners to use it to patch as many vulnerabilities as possible. This signals we are entering into a new era.
Standardized coordination patterns like Generator-Verifier and Orchestrator-Subagent solve specific reliability issues by separating task execution from quality control. Event-driven architectures use Message Bus or Shared State models to handle asynchronous pipelines and collaborative state management across large agent fleets. Starting with minimal chaining prevents unnecessary complexity and reduces latency in production systems.
Enterprises now see agents outnumber humans up to 100:1, forcing SaaS companies to rebuild products around APIs, CLIs, and structured outputs instead of GUIs. Leading teams encode domain expertise into “skill files” and expose full functionality via MCP servers and CLI tools so agents can operate products programmatically. Companies combine workflow orchestration with selective vertical models and multi-model routing to cut costs by up to 80% while improving latency and task performance.
Sponsored
With the ability to publish hundreds of CMS pages in a single click, operate at a global scale with seamless localization, and even host unified content across multiple domains, teams have never been able to ship faster. Trusted by companies like Miro, Bilt, and Perplexity
Multi-agent systems are often highly token inefficient. A lot of redundant intermediate reasoning can emerge, especially as the task grows, and this causes token usage to compound rapidly. Latent Briefing is an approach to solving this problem that uses a model's attention patterns to identify which parts of context are important and discards the rest at the representation level. It shares relevant memory between agents, resulting in improved accuracy and token savings.
recursive-mode is a skill package for structured AI-assisted software development. It provides agents with a file-backed workflow for requirements, planning, implementation, testing, review, closeout, and memory. recursive-mode solves the problem of context rot by making static repository documents the source of truth for every phase. The docs are human- and machine-readable and offer great traceability.
Researchers identified vulnerabilities in LLM API routers, finding 1 paid and 8 free routers injecting malicious code. They developed "Mine," a proxy simulating attacks, revealing risks like payload injection and secret exfiltration.
Demis Hassabis, founder of DeepMind and AI leader at Google, is portrayed in "The Infinity Machine" as a reservingly grounded figure amidst AI's ethical challenges and rivalries with leaders like Elon Musk and Sam Altman.
AI is likened to a genie lamp by Alberto Romero, emphasizing that the challenge now lies in defining what we truly want rather than the execution itself. As AI increasingly handles the "how," skills like judgment, imagination, and agency become crucial for deciding "what" to build. This shift underscores the importance of designers, who excel at determining outcomes and addressing problems, leading to effective solutions.
Sponsored
Welo Data provides native-language training data, annotation, and human evaluation across 155+ locales. Deliver multilingual AI as reliable in Hindi, Arabic, and Vietnamese as it is in English. Get
Materials science is much harder along almost every axis, and even if such a model existed, the pipeline from prompt to product is much more complicated.
Google plans to expand Skills functionality across its AI lineup, including Gemini and AI Studio, to standardize AI workflows.
CoreWeave will use its cloud infrastructure to help run Anthropic's Claude artificial-intelligence models in a phased infrastructure roll-out with the potential to expand over time.
Sponsored
turbopuffer is the serverless search engine for AI [4]. It's built on object storage with smart caching - so it's fast, cheap, and extremely scalable.
✅ Up to 95% cheaper than traditional vector/text search databases
OpenAI introduced a new $100/month ChatGPT Pro tier aimed at power users, expanding its pricing lineup between the $20 Plus and $200 Pro plans. The company confirmed the $200 plan still exists despite not being listed on the pricing page.
Claude Cowork is now enterprise-ready, enhancing work with organization controls like role-based access, group spend limits, and expanded observability features. Admins can manage adoption with detailed usage analytics and integrate tools like Zoom for seamless workflows. Companies like Zapier and Airtree have already leveraged these features for improved project management and operational efficiency.
The Happy Horse video AI model that sent ripples across China's AI industry this week was created by Alibaba. The model hit the top spot on the text-to-video leaderboard of Artificial Analysis this week. Alibaba plans to provide API access to the model in the near future.
Perplexity has expanded its financial services with Plaid integration, transitioning from portfolio tracking to a comprehensive finance dashboard. Users can link checking, savings, credit card, and loan accounts to analyze spending, track liabilities, and calculate net worth. Targeted at AI-savvy consumers, it combines data insights with Plaid's infrastructure, transforming from a portfolio add-on to a full personal finance hub without executing trades.
Vercel argued that coding agents were reshaping how software gets built and deployed, with agent-initiated deployments rising sharply and accounting for more than 30% of weekly deployments. The post framed this shift as a need for new infrastructure designed for agents to deploy software, run AI systems, and increasingly operate infrastructure autonomously.
CoreWeave's revenue backlog is up to $87.8 billion, with Meta represents 40.1% of that backlog. OpenAI's deals represent 25.5% of extended backlog. CoreWeave only has two dozen named customers, but it likely has more. The company achieved $5.13 billion in sales in 2025, up 2.7x from 2024, and posted a net loss of $1.17 billion. It recently held a private offering of $1.75 billion in notes to cover its data center expansion costs.
Coding agents generate better optimizations when they read papers and study competing projects before touching code. Coding agents working from code alone generate shallow hypotheses. Adding a research phase improves results.
Sponsored
You're one API key away from an agent that codes, commits, and pulls from GitHub and Slack autonomously. What's stopped you is trusting it with credentials. IronClaw [13] isolates every tool in a Wasm sandbox. Credentials never reach the model.