Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Sponsored
Rime's [10] AI voice models capture the tone, rhythm, and natural variability of real human speech. Designed by linguists, they're made for actual conversations in multiple languages and work on-prem and in the cloud. To try these low latency voices, sign up and get $100 in
Sponsored
61% of listeners chose Rime's AI voices over Google and ElevenLabs. Get $100 in free credits and start building a voice agent that
'Real-time RL' is a technique created by the Cursor team that involves using real inference tokens for training. The team serves model checkpoints to production, observes user responses, and aggregates those responses as reward signals. The approach allows them to ship an improved version of Composer as often as every five hours.
Chroma Context-1 is a 20B parameter agentic search model trained on over eight thousand synthetically generated tasks. It achieves retrieval performance comparable to frontier models at a fraction of the cost and with up to 10 times the inference speed. The model returns a ranked set of supporting documents to a downstream answering model, cleanly separating search from generation. It decomposes high-level queries into sub-queries and iteratively searches a corpus across multiple turns, selectively discarding irrelevant results to free capacity as its context window fills.
Cohere's Transcribe is an open-source ASR model optimized for low word error rate and production use. It achieved top accuracy on the Hugging Face Open ASR leaderboard.
Mistral's Voxtral TTS is a 4B-parameter multilingual text-to-speech model designed for low-latency, expressive voice generation, and scalable deployment in voice agents.
Anthropic has won a preliminary injunction in its lawsuit against the Department of War over the ruling of the company as a 'supply chain risk'. The judge in the case said that nothing in the governing statute supports the notion that an American company can be branded a potential adversary and saboteur of the US for expressing disagreement with the government. Anthropic refused to grant the Pentagon unfettered access to its models across all lawful purposes, as it didn't want its technology to be used for fully autonomous weapons or domestic mass surveillance. The company says that it will continue working productively with the government.
While AI companies guard their strategies closely, their hiring pages are public. These posts contain clues about what products companies are developing, who they hope to sell them to, and what bottlenecks they see coming. This post analyzes the current open roles at the leading foundation labs. As these labs continue to grow and their strategies change, their career pages will remain one of the best places to watch it happen.
Grok currently has the largest traffic share in the AI video generation space.
CapCut has begun rolling out Dreamina Seedance 2.0, a new video and audio generation model in CapCut with built-in safety safeguards, initially to paid users in select global markets.
Cline Kanban is a kanban board for managing coding agents that streamlines workflows by displaying task statuses and dependencies at a glance.
Sponsored
Charging per seat in SaaS once made sense, but not now, when your product charges per token, per API call, or per resolved ticket.
Google's TurboQuant is a compression algorithm that reduces the memory footprint of large language models while also boosting speed and maintaining accuracy. It reduces the size of the key-value cache so it doesn't have to be recomputed. Early testing shows TurboQuant results in an 8x performance increase and 6x reduction in memory usage without a loss of quality. Compression techniques like TurboQuant could improve the quality of outputs of models for edge devices without having to send data to the cloud.
ARC-AGI-3 was designed to evaluate agentic intelligence via interactive reasoning environments. Beating it will mean an AI system matches or exceeds human-level efficiency on all environments upon seeing them for the first time. 100% of the environments are solvable by humans on first contact with no prior training or instruction. All frontier AI reasoning models currently solve under 1%.
Reflection is a startup leading an effort to create freely available US AI systems. It is one of a handful of Nvidia-linked startups seeking to build a network of open source AI models. The startup is in talks to raise $2.5 billion at a valuation of $25 billion. Investors describe Reflection as the 'DeepSeek of the West' as it offers an alternative to the open source models offered by Chinese companies.
Manus' co-founders, Xiao Hong and Ji Yichao, have been told not to leave China while authorities review the company's $2.5 billion sale to Meta. Early versions of Manus were created by engineers from a Chinese company. A Singapore-based entity then took over Manus' operations and relocated most of its China-based employees to Singapore, which made it possible for Meta to purchase it. Authorities are concerned that Manus' moves could encourage other Chinese companies to follow suit and move out of the country without vetting.
Open source models are reaching parity with frontier labs' models, making those labs' equity look overpriced if they're simply utilities. These frontier labs have enterprise agreements, safety certifications, distribution, research talent, and regulatory positioning, but that doesn't explain their moat. People focus on capability, but the number that actually matters for valuations is the monetizable spread, the subset of that capability delta that someone will actually pay a premium for. The monetizable spread is declining faster than the capability spread.
Quantized models are actually pretty good. A 16-bit to 8-bit quantization carries almost no quality penalty - the difference with a 4-bit quantization is more noticeable, but it would only perform about 90% as well as the original. It's worth experimenting with these models as they're much smaller and can run on more systems. This article explains how model parameters work, what quantization is, how quantization is applied in practice, and the effects of quantization on model accuracy.
The final training run for a model is only the last step in a long, expensive process. Before the run, companies burn compute on running experiments at various scales, generating synthetic data, testing ideas, and training unreleased models. The full cost of developing a model is much higher than the cost of the final training run of a frontier model. Most of the spend is on exploration rather than execution. Companies that learn from their competition can replicate their results for a fraction of the original cost.
Sponsored
How can enterprises start using the massive amounts of data they're sitting on? Read the special report by 451 Research [13] to discover the latest innovations in RAG and beyond. When you're ready to start building, explore enterprise search with OpenSearch [13] - the trusted, open source platform for AI-powered search, analytics, and vector database solutions.
OpenAI outlined the philosophy and structure behind its Model Spec, a framework defining desired model behavior, safety principles, and how systems should follow instructions and resolve conflicts.
Manthan Gupta built Auto-Inference-Optimiser to let an AI agent hill-climb on LLM inference speed while keeping quality fixed on Apple Silicon. Argmax sampling and simplifying inference code gave the largest throughput gains, while most tuning knobs and KV cache quantization hurt or had no effect. The project highlights that a disciplined, observable harness is critical for distinguishing real performance wins from noise or benchmark illusions.
Cognition's Devin is an AI software engineer that can build software from start to finish without human involvement. When it launched in 2024, it was considered a step toward a long-held Silicon Valley dream of a machine that codes for you. Cognition's CEO, Scott Wu, believes that the technology doesn't mean the end of software engineering. Rather than eliminate engineers, Cognition's tools will allow them to focus on the best parts of the job while sparing them from the grunt work that traditionally consumes most of their time.
OpenAI launched a public bug bounty program targeting AI misuse and safety risks, expanding beyond traditional security vulnerabilities to include abuse scenarios.
Sponsored
Granola takes your raw call notes, enriches them, and makes them instantly valuable. Share with your team or prompt it to draft follow
AI companies are shifting from narrow solutions to broad platforms, driven by rapid model changes.
Harvey raised $200 million in a new round led by GIC and Sequoia, bringing its valuation to $11 billion and total funding past $1 billion.
Lyria 3 Pro extends the maximum track length to three minutes and adds finer control over song structure and customization.
Sponsored
You've built the perfect fine-tuned model. Now scale it without compromise. Generic cloud infrastructure wasn't designed for proprietary architectures, leading to performance bottlenecks. Crusoe's Bring Your Own Model (BYOM) changes that. Powered by MemoryAlloy™ technology, Crusoe optimizes its inference engine specifically for your unique weights.
Anthropic released Auto Mode in research preview, enabling Claude to autonomously execute actions with built-in safeguards that filter risky behavior and prompt injection.
OpenAI shifted away from its in-chat checkout feature after low adoption, prioritizing product discovery and merchant-directed purchasing flows instead.
Anthropic's Prithvi Rajasekaran developed a multi-agent architecture to improve AI-driven frontend design and full-stack application coding, addressing issues of coherence and self-evaluation. Inspired by GANs, this approach uses planner, generator, and evaluator agents to produce complex, high-quality outputs by decomposing tasks and utilizing structured handoffs. Despite improvements, challenges remain in context management and evaluator tuning, highlighting the ongoing need for adapting harness designs as AI models advance.
The App Store was a centralized answer to the distribution problem of a new computing platform. The agent era will need a new solution as agents need APIs, not app stores. Apple gained its revenue by forcing every in-app transaction through its payment system. The agent era lacks Apple's lock-in mechanics, so if one platform tries to charge high payment fees, users will just switch to a competitor. This suggests the payment layer will be competitive and low-margin rather than monopolistic.
As of March, Claude 4.6 features a 1M token context window and four distinct modes: Chat, Cowork, Code, and Projects. The Cowork suite automates workflows via Scheduled Tasks and Connectors, while the Code environment utilizes CLAUDE.md hierarchy, MCP protocols, and Agent Teams for autonomous development. Key upgrades include Computer Use research previews and deterministic Hooks for programmable guardrails.
Sponsored
First impressions matter. With Framer, early-stage founders can launch a beautiful, production-ready site in hours. No dev team, no code, no hassle. Join hundreds of YC-backed startups [11] who launched here and never looked back. Put your business in the perfect frame. Save $360 with a full year of Framer Pro [11].
Many of the modern workloads that LLMs are increasingly utilized for prioritize throughput over per-request latency, which many LLM systems and deployments optimize for today. Ray Data LLM is a library built for large-scale batch inference for LLMs. It provides scalable execution, high throughput, and fault tolerance. It has a highly optimized architecture for running LLM batch inference. Users can achieve 2x throughput with Ray Data LLM over vLLM's synchronous LLM engine while benefiting from production-scale resiliency.
Ossature is an open-source harness for spec-driven code generation. Developers write specifications describing what their software should do, and Ossature validates them, has an LLM audit them for ambiguities and gaps, produces an editable plan, and then generates code one task at a time. Each task only gets the context it needs. Ossature has verification built into the build loop. If verification fails, a fixer agent gets the error output and tries to repair the code.
Directional updates in RLVR were shown to better identify reasoning-critical tokens, enabling both test-time extrapolation and training-time reweighting to boost accuracy.
TurboQuant is a quantization method that reduces vector memory overhead while preserving performance. This improves key-value cache efficiency and accelerates vector search.
Semantic calibration appears to emerge as a byproduct of next-token prediction. Base models are remarkably well-calibrated when using a certain sampling-based notion of semantic calibration. They can meaningfully assess confidence in open-domain question-answering tasks despite not being explicitly trained to do so.
OpenAI has announced a new $10 billion commitment from a16z, DE Shaw Ventures, MGX, TPG, and T Rowe Price. The fresh capital brings OpenAI's record fundraise to over $120 billion. OpenAI has moderated its spending plans and is now targeting approximately $600 billion in total compute spend through 2030. It is now taking steps to prioritize its most profitable initiatives ahead of an IPO.
US District Judge Rita F. Lin of the Northern District of California said during a court hearing that the US government appeared to be punishing Anthropic by banning the company. The hearing is part of Anthropic's efforts to ease the government ban on the use of the company's AI models. Lin has yet to rule on the matter but expressed serious doubts about the Trump administration's actions in her opening remarks. The government's action has already cost Anthropic hundreds of millions of dollars in canceled contracts and aborted customer agreements.
Sponsored
Replay (San Francisco, May 5-7) is Temporal's practical conference for building prod-ready AI agents. Join the full experience or add on