Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
/goal is a headline feature that shipped in Codex on April 30 that introduced persisted goals, goal states that survive terminal restarts, laptop sleeps, and multi-hour passes without re-prompting. The feature injects a developer message on resume rather than waiting for the user to time anything. It allows users to turn off their devices and pick up exactly where they left off without re-prompting anything.
There needs to be a higher bar for what good quality looks like. Any vendor selling into a frontier lab is being judged implicitly during the purchase decision, and most are failing multiple quality control gates at once. Standardizing QC for evaluating how well data tests something on the performance-cost-latency Pareto curve is of the utmost importance. The vendors who don't internalize a higher QC bar will start running into problems this year.
Sponsored
73% of enterprises say poor data connectivity is preventing them from scaling AI. In this live webinar [11], Microsoft and CData walk you through the infrastructure that clears the bottleneck: universal data connectivity. Join on May 13th for an architecture blueprint for your
AlphaEvolve is a Gemini-powered coding agent that can design advanced algorithms. It can help make new discoveries on open problems across mathematics and computer science. The agent has been upgraded with the ability to help explain the physics of the natural world to help accelerate progress for scientists and businesses across a variety of fields. This post shares the areas AlphaEvolve has had the most significant impacts on to date.
Ramp Sheets built Fast Ask to handle its spreadsheet agent's information retrieval loop. It can navigate a workbook, read the relevant ranges, and return a compact answer for the main agent to use. This post presents Fast Ask as a case study for reinforcement learning. It covers where it is worth training a specialized agent, how to design the environment, and how to evaluate whether post-training worked.
ds4.c is an intentionally narrow, small native inference engine for DeepSeek V4 Flash. The project aimed to create one local model that felt finished from end to end. It is Metal-only, but the development team may implement CUDA support in the future. The project is still in alpha, so there are still bugs in the implementation.
Meta detailed In-Kernel Broadcast Optimization (IKBO), a co-design approach that removes redundant embedding replication in recommendation inference workloads.
Anthropic introduces Natural Language Autoencoders (NLAs) to translate AI model activations into human-readable text, aiding in understanding model thoughts. NLAs have been used to detect safety concerns and hidden motivations in AI behavior, improving model alignment auditing. Despite limitations like hallucinations and high costs, NLAs advance AI auditing techniques. Anthropic has released training resources for further development.
Chinese and American labs look similar when you look at their outputs and ingredients, but there are major differences in how they are organized and conditioned. Chinese scientists are much more willing to do non-flashy work to improve a model rather than push their own ideas. This results in less gamification of the system and more flexibility in adapting modern techniques. The Chinese AI community feels more like an ecosystem than battling tribes, with all the labs having nothing but respect for their peers. They are also much less focused on the business side of the technology.
Silicon Valley's narrative emphasizes AGI as the ultimate scarce resource, but the rapid commoditization of AI models challenges this. Intelligence now follows the same path as compute, bandwidth, and storage, where market forces drive competition and reduce costs. The real winners in AI won't necessarily have superior models but will own customer relationships and proprietary data, much like past tech giants.
Sponsored
The most powerful model can be held back by bad execution, orchestration, and observability. See how real teams are getting it
EVE Online is a uniquely rich environment for study, especially for developing AI systems that use long-horizon planning, memory, and continual learning.
TLDR's Applied AI team is tasked with making every process at TLDR legible to code, runnable by anyone, and composable into larger workflows. Join a small, fast moving team using the latest AI tools with an unlimited token budget. Learn more [21].
OpenAI introduced Trusted Contact, an optional feature that allows adults to nominate someone who may be alerted if severe self-harm risk is detected in conversations.
Perplexity released Personal Computer for all Mac users through its desktop app, giving AI agents access to local files, applications, connectors, and the web.
Sponsored
How much AI are your customers ready for? Well, 86% of AI users say it helps them better understand their finances. Consumers aren't afraid.
Anthropic increased usage limits for Claude through a new compute partnership with SpaceX, accessing over 220,000 NVIDIA GPUs. This expansion follows deals with Amazon, Google, Broadcom, Microsoft, NVIDIA, and Fluidstack for significant compute capacity. The company also plans international expansion to address compliance needs for enterprise customers in regulated industries.
Claude Managed Agents launched features like dreaming, outcomes, and multiagent orchestration. Dreaming enhances agent improvement by analyzing past sessions to identify patterns, while outcomes allow agents to self-correct based on predefined success criteria. Multiagent orchestration optimizes complex task management by enabling agents to delegate tasks to specialized subagents, as utilized by companies like Harvey, Netflix, Spiral by Every, and Wisedocs.
DeepSeek is in talks to raise money from China's National Artificial Intelligence Industry Investment Fund, a one-year-old government-backed fund with around $8.8 billion in capital. The startup aims to raise a few billion dollars in the new round, which values it at around $50 billion. DeepSeek is a key component in China's plan to have top-class homegrown companies in a range of AI fields. The strategy is a way to hedge against US export controls and to take leadership in bringing AI to the world.
OpenAI's Codex now surpasses Anthropic's Claude Code after Codex's integration of GPT-5.5 and improved app performance. Austin Tedesco highlights Codex's use in creating strategy documents from diverse sources, while Dan Shipper uses it for recruiting based on career trajectories. Marcus Moretti adopts a cautious approach to new AI tech, focusing only on tools solving real problems and proven by reputable use.
Language language models forget everything the moment they finish replying. Memory systems help them 'remember' things so they can have conversations. Agent memory systems are a part of the loop that carries information forward. This article looks at different ideas on what information should be passed on in each loop.
Sponsored
General-purpose AI agents fail in specialized domains — subtly wrong in edge cases. Domain specialization fixes this. Build AI agents with four levers [11]: system prompt, knowledge corpus, tool selection, guardrails. Demonstrated across customer engagement,
Multipath Reliable Connection (MRC) is an RDMA transport protocol that enables a single RDMA connection to distribute traffic across multiple network paths. This improves throughput, load balancing, and availability for large-scale AI training fabrics. MRC delivers high levels of GPU utilization by load-balancing traffic across all available paths. It gives administrators fine-grained visibility and control over traffic paths to simplify operations and accelerate troubleshooting at scale.
TokenSpeed, a high-performance LLM inference engine, optimizes agentic workloads with speed-of-light efficiency, leveraging a compiler-backed modeling mechanism and a high-performance scheduler. It delivers faster throughput than TensorRT-LLM for coding agents, with optimizations like TokenSpeed MLA to enhance Nvidia Blackwell's performance. Developed with NVIDIA DevTech and other collaborators, TokenSpeed significantly reduces latency and increases throughput in typical agentic workloads.
The vLLM V1 update improved inference correctness by addressing discrepancies in logprob computation, runtime defaults, inflight weight updates, and final projection precision. Key fixes included adjusting processed logprobs, disabling prefix caching, matching weight update models, and ensuring fp32 lm_head computation to align with vLLM V0's behavior. These changes resolved initial training mismatches, ensuring the new engine maintains expected RL performance without unnecessary objective-side corrections.
ProgramBench challenges agents to recreate software executables without source code, using only documentation and experimentation. The tasks range from terminal utilities to complex software like compilers and libraries, offering over 248,000 behavioral tests across 200 tasks. Agents must design and implement entirely from scratch in a secure, sandboxed environment, emphasizing software architecture skills without external aids or decompilation.
Google is betting that enterprise AI is a platform problem, not a services problem. It is in talks with Blackstone, KKR, and EQT to give their portfolio companies access to Gemini models through omnibus licensing agreements. The discussions are not exclusive, and no deals have been finalized. Google is offering private equity firms a commercial wrapper that gives their entire portfolio access to Gemini, then relying on the consulting ecosystem it has already financed to handle implementation. The approach trades consulting revenue for distribution speed.
AI inference demands extreme data performance, overwhelming traditional storage and data infrastructures. Vector DBs, sub-millisecond access times, and decoupled cloud storage are essential to handle unprecedented concurrency and unpredictable workloads. Silk offers a solution that boosts storage performance without heavy provisioning, keeping systems resilient against AI-driven demand spikes.
World models aim to advance AI from mere pattern recognition to understanding and interacting with the physical world, posing potential challenges like data friction and variation. Investments from AI pioneers like Yann LeCun are addressing these obstacles with significant billions to develop models that encapsulate complex physical interactions beyond current LLM capabilities. The struggle remains in obtaining diverse, high-quality, real-world data necessary for these models to function effectively, creating a significant challenge and opportunity in AI progression.
Sponsored
turbopuffer wrote about building a 100B-vector search index. The post examines turbopuffer's architecture, travels up the modern memory hierarchy, zooms into a single CPU core, and backs out to the scale of a distributed cluster. Read the blog [19].
Frontier model training depends on reliable supercomputer networks that can quickly move data between GPUs.
Sometimes, stable, self-reinforcing behavioral states emerge in large language models that resist suppression and sometimes spread into contexts far removed from the ones that produced them.
Systems keep getting better, and theorems keep arriving to explain why they can not - both can be true because they're usually about different things.
The design of subscription plans is being challenged by evolving product capabilities and usage patterns.
Moonshot has more than quadrupled its valuation in the span of just a few months.
Harvey's Legal Agent Benchmark (LAB) is an open-source tool for assessing AI agents' performance in legal tasks.
Google is testing screen sharing and custom agents in its Antigravity IDE.
Sponsored
OpenAI released GPT5.5 Instant, updating its default ChatGPT model with improved factual accuracy, reduced hallucinations, and stronger personalization based on user context.
Subquadratic has launched a new AI model with a 12-million-token context window. It outperforms GPT-5.5 on retrieval benchmarks. Attention cost scales quadratically with context length, so doubling the input quadruples the work. Subquadratic claims to have solved the problem. It plans to offer a model with a 50-million-token context window soon.
Meta is building a highly personalized AI assistant that will be able to carry out everyday tasks. The digital assistant will be powered by the company's new Muse Spark AI model. It can connect several hardware and software tools and learn from data with less human intervention than a chatbot. Meta is targeting a launch before the fourth quarter of this year.
A lot of LLM inference is transferring data from one place to another and then computing on it when it's there. The most frustrating bottleneck in the system is when compute units sit idle because the data bus feeding them isn't fast enough. The solution is to transform memory into compute. Quantization is a nice trick, but it doesn't actually trade memory for compute - it transfers half as much data to a place to do twice as much computation.
Vision agents are the default for operating web apps that don't expose APIs. Most teams default to vision agents because the alternative, writing an MCP or REST surface, is too expensive to build. The cost of the vision approach is treated as a fixed price. Current vision agents require detailed prompts to succeed in tasks, and they are still prone to making mistakes. Better vision models reduce error rates, but they do not reduce the number of screenshots required to reach the relevant data, each of which is worth thousands of input tokens.
Sponsored
Does your AI know how people convey tone, humor, and feelings in their mother tongue, or does it just translate from English? Welo Data's native-language training data [12] & human evaluation lets you build for your users, everywhere. Surface multilingual quality and