Mercor Raises $750M Series D at $20B Valuation, Hits $3B Run Rate
Mercor Raises $750M Series D at $20B Valuation, Hits $3B Run Rate
AI recruiting platform Mercor closed a $750 million Series D at a $20 billion valuation, with a $3 billion annualized run rate. The round signals continued investor appetite for AI-native hiring infrastructure, and the run rate indicates sustained enterprise demand. The same day, Google launched Gemini 4 Carbon in general availability.
AlphaGenome 2 Predicts How DNA Variants Affect Specific Cells and Disease
Google DeepMind released AlphaGenome 2, a model that predicts how DNA sequences and their variants impact specific human cell types and disease. The model deciphers how cells read DNA, linking genetic variants to concrete disease outcomes. For genetics research, this enables faster prioritization of disease-relevant variants and reduces reliance on wet-lab experiments.
Anthropic IPO Roadshow Video Reveals Compute Strategy Details
Anthropic kicked off its IPO roadshow yesterday, and the accompanying video contains previously unreleased details about how the company’s business operates. The section on ‘Approach to compute,’ featuring co-founder Tom Brown, is particularly notable, describing how Anthropic plans and secures compute resources. The material offers insight into the company’s cost structure and business model ahead of its public listing.
AWS Releases Official Local Emulator with 29,684 Commits in Three Weeks
AWS released aws-emulator, an official emulator that lets developers run local clones of its services for testing without provisioning a paid API key. Written in Rust and built by a large swarm of AI agents, the project logged 29,684 commits in its first three weeks. For developers, local testing of AWS services lowers cost and speeds iteration without consuming real cloud resources.
Google Launches Gemini Decisions API, Directly Rivaling Jev
Google released the Gemini Decisions API, powered by the new Gemini Decision 1 model, entering the decision-model category alongside Jev and OpenAI. The API supports three question types: yes/no, multiple choice, and rubric scoring—reusing Jev’s term ‘Noul’ for yes/no queries. The interface closely mirrors Jev’s design, giving developers a Google-native option for structured decision tasks without third-party dependencies.
Study Benchmarks Cost-Success Trade-offs Across Agent Architectures on Enterprise SOPs
A new paper presents a controlled evaluation of competing LLM agent architectures on enterprise-style standard operating procedures (SOPs): long, branching, tool-intensive workflows. By holding conditions constant, the study compares cost and task success rates across architectures, filling a gap in controlled evidence that enterprises previously lacked when choosing an agentic design. The results give teams deploying agents for business processes a concrete basis for weighing success rates against compute costs.
New Tokenizer Covers 204 Languages and Uses Fewer Tokens Than GPT-4o's, Even on English
Researchers trained tokenizers from scratch on human-written text in roughly 200 languages, producing Havon v3. On GPT-4o’s tokenizer, Indian languages require 1.6 to 14 times as many tokens as English for the same content, driving up serving cost and latency. Havon v3 covers 204 languages and needs fewer tokens than GPT-4o’s tokenizer, including on English. The result points to lower inference costs for multilingual applications, particularly for Indian-language use cases that current tokenizers handle inefficiently.
Developer's AI Agents Discover an Unmapped Maya Site
Developer Pietro Miotti writes that his AI agents found a previously unmapped Maya site. The post drew 914 points and 335 comments on Hacker News. It describes the role the AI agents played in the discovery and sparked broad community discussion about AI’s use in archaeological survey work.
United Airlines Bans AI Agents from Booking Flights for Travelers
United Airlines has banned AI agents from booking flights on behalf of travelers. The story drew 296 points and 272 comments on Hacker News. The policy directly affects AI assistants and travel-agent tools that rely on automated booking, forcing developers to adjust product logic or find alternative channels. It signals tightening boundaries for AI agents in regulated ticketing scenarios.
Real-Time Error Correction Catches LLM Reasoning Mistakes Mid-Step
A new paper addresses a common LLM failure mode: confidently written but hard-to-spot errors in math, science and engineering work, where a single wrong step propagates and wastes inference tokens. The proposed approach performs real-time error correction during reasoning rather than post-hoc checking, aiming to catch mistakes where they happen and reduce token waste from error accumulation.
TimeOmni-o1 Couples Time Series Foundation Models with LLMs for Forecasting
A new arXiv paper introduces TimeOmni-o1, which bidirectionally couples time series foundation models with LLMs for multimodal reasoning and forecasting. Real-world time series often come with accompanying text, such as reports on events affecting them. The model handles both question answering about a series and future value prediction, with bidirectional coupling letting text context and series data inform each other rather than being simply concatenated.
Physics2Weight Uses LLMs to Generate Neural Network Architectures and Weights from Physical Laws
A new paper extends weight-space programming (WSP) to physical problems. WSP generates neural network weights directly without gradient-based optimization, but applying it to physics has been challenging. Physics2Weight uses large language models to generate both network architecture and weights directly from physical laws. The work explores a new use of LLMs in building scientific computing models.
ADIndex Cuts Long-Context Inference Latency with Layer-Adaptive Sparse Attention Indexing
A new paper introduces ADIndex, a method targeting a bottleneck in DeepSeek Sparse Attention (DSA), where the indexer still scores the entire causal prefix for each query, making indexing a dominant source of latency in long-context inference. ADIndex uses group-level distillation to achieve layer-adaptive sparse attention indexing, reducing indexing compute. The work falls under long-context inference acceleration, aiming to cut the indexing share of latency.
Some Open Source Projects Begin Banning AI-Generated Contributions
A growing group of open source projects is explicitly rejecting AI-generated code contributions, collected on the site osssaysno.com. The topic drew 61 points and 106 comments on Hacker News. The debate centers on copyright ownership, quality accountability, and the review burden placed on maintainers. For developers using AI coding tools, checking a project’s policy on AI-generated contributions before submitting code is now necessary.
HN Debate: Retirement Savings Are Fueling the AI Bubble
A Hacker News thread with 130 points and 140 comments argues that ordinary people’s retirement savings are flowing into AI-related assets through investment channels, helping fuel the current AI valuation bubble. The discussion focuses on the link between pension allocations and the AI investment boom, and what this funding structure means for retail investors.
OpenAI Showcases Voice-Steered Codex and Cloud Handoff at AIE Code Summit
On day one of the AI Engineer Code Summit, OpenAI demonstrated new ways for developers to work with Codex: steering it by voice, handing off work to the cloud, and using dots. The capabilities point toward developers expressing intent and managing asynchronous tasks rather than writing code line by line. For teams using Codex daily, voice control and cloud handoff change how work is delegated and coordinated.
Gilbert + Tobin Rethinks Legal Business with AI
OpenAI published a conversation with Sam Nickless, Partner and CEO of Australian law firm Gilbert + Tobin, on how the firm is integrating AI into its legal practice. The discussion covers changes to client service, internal workflows, and business model. It offers a practitioner-level view of AI adoption in law rather than abstract commentary.
Google Labs Launches Tabletop for Group Planning with Itineraries, Polls and Brackets
Google Labs has launched Tabletop, a new experiment for group collaboration. Users can build interactive itineraries, polls and brackets to plan activities with their team. The product is currently invite-only via a waitlist. Tabletop is positioned as a shared space for group planning and decision-making, suited to use cases like trip planning, event organization or small tournaments.
HN Discussion: AI, Activity and Outcomes in Software Engineering
A Hacker News thread linking to a YouTube video drew 103 points and 99 comments on the relationship between AI use, developer activity and outcomes in software engineering. The discussion centers on how AI tools affect engineering productivity.
Simon Willison Builds Tag-Based Browser for His 244 Vibe-Coded Tools
Simon Willison used leftover Fable 5.2 credits to build a tag-based browsing page for his collection of 244 vibe-coded tools at tools.simonwillison.net. He used tree-sitter to analyze the code for JavaScript APIs and libraries. The goal was a better way to navigate the growing collection, and tags now let users locate tools quickly. For developers, it’s a practical example of using code analysis to auto-organize a personal tool library.