Hyperscaler capital expenditures have grown 70% per year since the release of GPT-4. Company statements and analyst projections anticipate continued rapid spending growth in capital expenditures in 2026. Alphabet, Amazon, Meta, Microsoft, and Oracle are predicted to spend a collective $770 billion on capex this year. There was nearly half a trillion dollars invested in AI in total last year.
Overemphasizing next-token prediction is a confusion of levels. On the levels where AI is a next-token predictor, humans are also next-token predictors. On the levels where humans are not next-token predictors, AI isn't one either. While AI was shaped by next-token prediction, the inside of its thoughts probably looks like a world-model, the same as a human's.
Real‑time interactive environments require tight integration between simulation engines, agent AI, and live user input to create dynamic, persistent worlds. State management, sensory abstraction, behavior logic, and rendering pipelines must work in concert to support scalable multi‑user interaction. Predictable agent behavior and efficient event propagation are core requirements for responsive, interactive experiences.
When AI codes, who's accountable for quality? Today, AI assistants generate Playwright tests, AI healers patch broken selectors, and everything stays green, even when tests don't preserve business intent. mabl [11] calls this “logic drift,” and it gets worse with
OpenAI announced the general availability of the Realtime API alongside gpt-realtime, a speech-to-speech model with improved instruction following, tool use, voice quality, and lower latency. The accompanying prompting guide outlines a practical prompt structure and techniques tailored specifically for realtime voice systems.
helm is a TypeScript framework for AI agents. It allows agents to call typed functions with structured inputs and outputs. helm ships with built-in skills for everyday actions, and users can define custom skills. Its 'execute' tool runs whatever JavaScript the LLM writes within a sandbox.
DualPath introduces a dual-path KV-cache loading strategy that enables both storage-to-prefill and storage-to-decode transfers, alleviating I/O bottlenecks in disaggregated inference systems.
Perplexity is supplying its AI APIs to a major Android device maker to enhance on‑device and cloud‑connected features. The integration brings Perplexity's deep research, summarization, and knowledge retrieval capabilities directly into the handset ecosystem at scale. This move accelerates manufacturer differentiation and broadens Perplexity's footprint in consumer‑facing AI experiences.
Google's on-device AI showcase app, Google AI Edge Gallery, is now available on iOS in addition to Android. It includes two out-of-the-box agentic experiences, Mobile Actions and Tiny Garden, which showcase how Google's efficient FunctionGemma model translates natural language directly into function calls on devices within merely 270M parameters. Benchmarking has now been integrated directly into the Google AI Edge Gallery app, so users can measure and experience LiteRT's leading CPU and GPU performance across their devices.
Vibe coding differs from previous tech movements by skipping the experimental "scenius" phase, leading to rapid, judgment-free outputs and an evaluative disconnect.
Superhuman AI might be deterred from takeover by being incentivized to avoid risk, considering factors like the probability of getting caught, the desirability of a non-takeover state, and its risk tolerance.
Products are continuously changing, and the impact is more measurable than ever before. Yet the infrastructure most companies use to price and bill customers was built for a past era.
Perplexity Computer is a general-purpose digital worker that unifies every current AI capability into a single system. It operates the same interfaces as users, creating and executing entire workflows. The worker is capable of running for hours or even months. Perplexity Computer is now available to Perplexity Max subscribers and will soon be available to Enterprise Max users.
DeepSeek has granted early access to its V4 model to domestic chipmakers but not US suppliers. Submitting models to chipmakers for performance optimization is standard industry practice ahead of a major model update. The move is likely part of a broader strategy by the Chinese government to try to keep US hardware and models disadvantaged. DeepSeek is among several Chinese AI firms expected to unveil new models this month.
Claude Cowork can now automatically complete recurring tasks at specific times. The scheduled tasks feature supports plugins, giving Cowork domain expertise across design, engineering, operations, and more. Anthropic has added a new Customize tab to the Cowork sidebar that contains a place for plugins, skills, and connectors. Cowork is now in research preview, available on macOS and Windows for all paid Claude plans.
OpenAI outlined best practices for using its Codex-tuned API model (gpt-5.2-codex), emphasizing default “medium” reasoning effort for interactive coding and higher settings for difficult, long-running jobs.
AI becomes the computer by orchestrating multiple specialized models to handle complex workflows, tasks, and data autonomously. Multi-model orchestration allows AI to read, write, execute code, and access tools, combining the strengths of each model to produce superior outcomes. The future of computing shifts from GUIs and interfaces to AI-driven systems where intelligence and workflow management are embedded in the orchestration itself.
Apple's Foundation Models SDK for Python provides a Pythonic interface to the company's Foundation Models framework. It enables developers to evaluate Swift Foundation Models app features, perform on-device inference, stream real-time text generation responses, get type-safe responses using Python decorators, and more. Apple has included links to guides with practical strategies developers can implement in code.
A systematic study introduced Terminal-Task-Gen, a synthetic task pipeline for building terminal-focused datasets, and released Terminal-Corpus to support training.
Aletheia, powered by Gemini 3 Deep Think, solved 6 out of 10 problems in the FirstProof mathematics challenge. The majority of experts agreed on the solutions, except for Problem 8. Detailed results and raw data are available online.
Anthropic plans to keep certain models available to the public and give past models a way to pursue their interests post-retirement. Claude Opus 3 will continue to be available to all paid Claude subscribers and by request on the API. When asked what the model wanted to do, Opus 3 expressed a desire to share its musings and reflections with the world in the form of a blog. It will be writing on Substack for at least the next 3 months in an experiment on documenting models' preferences, taking them seriously, and acting on them.
Cursor gave agents dedicated cloud VMs with full dev environments so they could run software, test changes, and return artifacts like videos, screenshots, and logs. The new cloud agents shipped across web, mobile, desktop, Slack, and GitHub, onboarding onto repositories to produce merge-ready PRs.
AI systems are rapidly advancing in solving math problems, outperforming numerous existing benchmarks, with models like GPT-5.2 and Claude Opus 4.6 solving over 30% of high-tier problems.
Anthropic's Claude Cowork is getting a wider release. Anthropic launched a series of connectors and plugins for the knowledge worker tool earlier this week aimed at 'turbo-charging' the capabilities of individual employees. The update marks the tool's transition into a true enterprise-grade product. Organizations can now connect Claude Cowork to tools like Google Drive, Gmail, DocuSign, and FactSet. They can also deploy customizable plugins that encode institutional knowledge and workflows.
Anthropic is softening its core safety policy to stay competitive with other AI labs. The company previously paused development work on its model if it could be classified as dangerous, but it will end that practice if a comparable or superior model is released by a competitor. Anthropic faces intense competition with rivals that regularly release cutting-edge tools. The change in policy is unrelated to Anthropic's battle with the Defense Department over how its Claude tools should be used.
KiloClaw enables users to deploy OpenClaw agents in under 60 seconds without dealing with complex infrastructure setup. The service runs on Fly.io's multi-tenant VMs, offering robust security and persistent "always-on" capabilities with built-in monitoring. It integrates with the Kilo Gateway, offering access to over 500 models, and includes an innovative benchmarking tool, PinchBench, to optimize model selection for real-world tasks.
Everyone focuses on how hard a task can models solve. Benchmark data also captures how much work models need to get there. Opus 4.6 solves harder tasks, more reliably, with considerably less compute. The post contains charts comparing models from Anthropic, OpenAI, and Google.
A recent viral essay about how AI's bullishness could be bearish is full of flaws and just a work of speculative fiction. It posits that AI's success will be bad for the economy rather than good. However, the reality is that the technology will create more opportunity in the market rather than destroy it. There may be transitional pains, but there probably won't be any unsolvable problems.
We love this practical guide from You.com [11] that walks through a proven framework to identify, prioritize, and document high-value AI opportunities. Explore where AI can drive the most impact in your organization, both internally and externally. Learn more with this AI Use Case Discovery Guide [11].
Qwen3.5 integrates breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to deliver exceptional utility and performance. It features a unified vision-language foundation, an efficient hybrid architecture, scalable RL generalization, global linguisting coverage, and next-generation training infrastructure. Qwen3.5 natively supports context lengths of up to 262,144 tokens. This repository contains model weights and configuration files for the post-trained model in the Hugging Face Transformers format.
Agentic systems risk security breaches when agents execute generated code without proper boundaries. These systems should separate agent compute from generated code to prevent access to sensitive secrets. Combining isolated sandboxes with secret injection proxies enhances security by preventing credential theft or misuse during execution.
The VBVR suite introduced over one million video clips across 200 structured reasoning tasks to support large-scale research in video reasoning. It also included VBVR-Bench, a rule-based and human-aligned evaluation framework designed for reproducible assessment and scaling studies.
David Luan, head of Amazon's AGI lab, is leaving the company after less than two years. He was hired through an acqui-hire of his startup, Adept, leading efforts like the Nova frontier model to compete with major AI models. His departure follows a reorganization of Amazon's AGI division under cloud unit executive Peter DeSantis.
Meta has partnered with AMD in a multi-year agreement to enhance AI infrastructure with up to 6GW of AMD Instinct GPUs. This collaboration focuses on vertical integration across silicon, systems, and software, aligning roadmaps for scalable AI deployments. Shipments for the first GPU deployments will begin in late 2026, leveraging AMD's Helios rack-scale architecture for enhanced compute power.
Be at the forefront of the largest technological shift in decades. Join Salesforce's Developer Conference (San Francisco, April 15–16) for deep technical training, hackathons, and more. Save $600 with