Perplexity Launches Computer for Taxes to Draft, Review, and Optimize U.S. Federal Tax Returns Using Loadable AI Tax Modules
Perplexity has expanded its Computer AI platform with dedicated tax modules that can automatically draft U.S. federal tax returns on official IRS forms, review professionally prepared returns for errors, and create custom tools for complex scenarios such as rental portfolio deductions or startup equity modeling. The modules leverage Perplexity's Agent Skills protocol and incorporate the latest tax law, including provisions from the OBBBA budget legislation that older software may miss. In internal testing, the system identified a 67% understatement of No Tax on Overtime deductions in an attorney‑prepared return. The launch aims at users who need more than off‑the‑shelf tax software but do not want to pay for full professional preparation.
Overwhelmed by Applications, Recruiters Turn to AI Pre-Screening Tools to Winnow Down Applications
Recruiters are drowning in applications as hiring slows, with 72% of global firms still reporting a talent shortage. Two‑thirds of talent‑acquisition professionals say they received more candidates per role last year, yet 46% say candidate quality has declined. In response,...
![[OFC 2026] Part 3 of 5: Packaging: Glass Substrates to 3D Hybrid Bonding](/cdn-cgi/image/width=1200,quality=75,format=auto,fit=cover/https://substackcdn.com/image/fetch/$s_!FG3y!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3201360a-f898-4593-8267-8cfc0766cc04_1280x1280.png)
[OFC 2026] Part 3 of 5: Packaging: Glass Substrates to 3D Hybrid Bonding
The OFC 2026 session highlighted that AI‑driven data centers are now limited by packaging rather than raw silicon speed. Co‑Packaged Optics (CPO) aims to overcome this bottleneck by integrating optical engines directly into switch packages, but the choice of substrate...
Intel Partners with Tesla and SpaceX on Terafab
Intel announced a partnership with Tesla, SpaceX, and xAI to create the Terafab joint venture, a massive vertically integrated semiconductor facility in Austin, Texas. The fab aims to deliver about 1 terawatt of AI compute per year, supporting billions of...
Inside View: Stewart Crane on Withers’ Technology Priorities and AI in Practice
In a recent Inside View interview, Stewart Crane, global IT director at Withers, outlined the firm’s technology agenda and its pragmatic AI rollout. Withers was an early adopter of Microsoft Copilot, deploying it firm‑wide to build internal AI expertise before...
The True Cost of Doing Nothing
The post argues that the real expense for B2B founders is not the $950 monthly fee for an AI prospecting system, but the revenue lost from missed client conversations. A typical deal worth $25,000 can disappear when prospecting is inconsistent,...
MCPs vs APIs in a Production Enrichment Pipeline
Rick Koleta’s GTM Vault episode shows how Skyp’s enrichment pipeline combines Claude Code’s plan mode with the Apollo API to deliver high‑quality leads at roughly fifty cents each. The build demonstrates that while MCP connectors (Gmail, Stripe, Grain, Slack) provide...
7 Steps to Mastering Retrieval-Augmented Generation
Retrieval‑augmented generation (RAG) enhances large language models by coupling them with external knowledge bases, addressing hallucinations and outdated information. The article outlines seven essential steps—from data selection and cleaning, through chunking, embedding, and vector‑store population, to query vectorization, context retrieval,...
Claude Code Pattern 6: Context Management at Scale
The Claude Code harness introduces a layered context‑management system to keep long‑running AI agents within the model's finite token window, typically around 200,000 tokens. It reserves up to 20,000 tokens for auto‑generated summaries and monitors usage with multiple thresholds that...
Virtual Stages, AI, and the Creator Economy: Cannes Film Market Massively Scales Up Its Innovation Hub
The Cannes Film Market (Marché du Film) is expanding its Innovation Hub for the May 12‑20 edition, adding virtual stages that let filmmakers pitch projects to global buyers without traveling. AI‑driven matchmaking and analytics tools will pair content with the most...
Anthropic's Annualized Revenue Tops OpenAI, Secures Major AI Chip Deal
Anthropic announced that its annualized revenue (ARR) now exceeds $30 billion, overtaking OpenAI’s $25 billion run rate. The surge represents roughly $21 billion of new revenue in just over three months, propelling ARR from about $9 billion at the end of 2025 to a...
Data Governance in the AI Era: 10 Shifts Redefining Data, Institutions, and Practice
The essay argues that data governance is the foundation of AI governance, as AI systems depend on high‑quality input data. It outlines ten transformative shifts, including redefined data definitions, expanded ownership, real‑time pipelines, and new ethical risk assessments. These changes...
April Job Market Update (the Data)
AI is rapidly displacing high‑skill white‑collar jobs, with 78,000 tech positions cut in the first half of 2025 alone. The Bureau of Labor Statistics reports the lowest professional‑services job‑opening rate since 2013, a 20% year‑over‑year decline, while routine‑cognitive roles fell...
Usage Limits Were Just the Beginning
In March 2026 Anthropic’s Claude suffered a cascade of failures: peak‑hour throttling, a prompt‑caching bug that inflated per‑message costs ten‑to‑twenty‑fold, the expiration of a temporary 2× usage promotion, and a series of platform outages. The combined effect drained usage meters...
The EU AI Act Newsletter #99: Bridging the Atlantic
The European Parliament approved a simplification package for the EU AI Act, setting firm deadlines of 2 December 2027 for high‑risk AI systems and 2 August 2028 for sector‑specific safety rules, while giving providers until 2 November 2026 to meet watermarking requirements. The package also bans...
6 AI Prompts Built From the Claude Code Leak
Anthropic’s Claude Code npm package unintentionally shipped a 59.8 MB source map that revealed 44 hidden feature flags and an advanced background agent suite. The leak exposed KAIROS, an always‑on repo monitor, /autoDream for memory reconciliation, and ULTRAPLAN capable of 30‑minute...
Let Claude Cowork Run Your Workflow: Practical Guide
Anthropic introduced Claude Cowork, a desktop‑mode AI that moves beyond answering questions to completing multi‑step knowledge‑work. By granting the model controlled access to local files, users can ask it to read, process, and output finished deliverables such as reports, spreadsheets,...
A Guide to Using Codex CLI and Some Best Practices
OpenAI’s Codex CLI, a Rust‑powered AI coding agent, is gaining traction among developers seeking deeper automation. The tool organizes its functionality into distinct layers—command, agent, skill, plugin, configuration, and hook—each governed by TOML files and slash commands. Best‑practice guidance now...
Tennant Rolls Out X16 SWEEP for Round-the-Clock Autonomous Sweeping
Tennant announced the X16 SWEEP, its first autonomous floor‑sweeping robot from a major cleaning brand, targeting warehouses, logistics hubs and light‑manufacturing sites. Powered by Brain Corp’s BrainOS, the machine combines AI‑driven path planning, visual intelligence and a DustShield system that...
Inside KeyBank’s Approach to Customer Experience and AI
KeyBank is reshaping its customer experience by embedding generative AI across its digital channels. The bank pledged roughly $200 million for AI projects, launching chat‑bots that now resolve about 40% of routine inquiries and a recommendation engine that lifted cross‑sell revenue...
AI Sleep Apnea Detection Tool Draws 150,000 Pre-Launch Sign-Ups
Sleep Cycle, the AI‑powered sleep‑tracking app, announced that more than 150,000 users have pre‑registered for its upcoming AI‑driven sleep apnea risk detection tool. The sign‑ups, mainly from the United States and United Kingdom, skew toward adults over 45, a group...
10 LLM Engineering Concepts Explained in 10 Minutes
The article reframes large language model (LLM) development as a systems problem rather than pure prompt engineering. It outlines ten core engineering concepts—including context engineering, tool calling, the Model Context Protocol, agent‑to‑agent communication, semantic caching, contextual compression, reranking, hybrid retrieval,...
QCNNs Classically Simulable Up To 1024 Qubits
Researchers led by Pablo Bermejo have shown that quantum convolutional neural networks (QCNNs) can be faithfully simulated on classical hardware for systems as large as 1,024 qubits. The study attributes the apparent quantum advantage to QCNNs processing only low‑bodyness, or local,...
Harvey Drives Legal Agent Learning Via ‘Harness Engineering’
Harvey conducted a small‑scale experiment using “harness engineering” combined with autoresearch to improve its legal AI agents. Across 12 benchmark tasks, the average success score jumped from 40.8% to 87.7%, with seven tasks surpassing 90% and one achieving 100% completion....
The Sequence Knowledge #838: Project GENIE: Building Playable Worlds From Pixels
Project GENIE, Google’s Generative Interactive Environment, moves AI beyond text‑only models toward real‑time video world simulation. By tokenizing raw pixels, the system builds a dynamic internal representation that reacts to user actions, effectively turning a viewer into an active participant....
FIRESIDE CHAT: Geopolitical Turmoil, Rising AI Risk Add a New Layer to Enterprise Cyber Defense
At RSAC 2026, enterprise security leaders highlighted a dual crisis: a surge of unsanctioned AI tools and growing geopolitical distrust of U.S.-controlled cloud services. Skyhigh Security’s Sanjay Castelino reported that European firms are reassessing reliance on American cloud infrastructure, demanding...
The Saudi Arabia of Lithium.
OpenAI released a comprehensive policy blueprint urging the U.S. government to tax, regulate, and redistribute wealth generated by advanced AI, framing it as a new social contract akin to the Progressive Era and New Deal. CEO Sam Altman warned that...
LawNext Podcast: Learned Hand’s Shlomo Klapper on Why Courts Are the Next Frontier for Legal AI
In a LawNext podcast, Shlomo Klapper, founder and CEO of Learned Hand, argues that courts represent the next major frontier for legal AI. Drawing on his experience as a litigator and appellate clerk, Klapper describes a new "reasoning engine" that...
How to Move Beyond AI Adoption to AI Depth
Chief People Officer Danny Guillory at Gametime has moved beyond counting AI adopters to measuring how deeply employees embed AI in daily work. Nearly 100% of staff now run AI agents for routine tasks, and Guillory’s own workflow includes AI‑drafted...
India’s Frugal AI Models Are a Blueprint for Resource-Strapped Nations
India is pioneering a frugal, sovereign AI strategy that emphasizes lightweight models capable of running on low‑end smartphones and low‑bandwidth networks. Initiatives such as AI4Bharat and startups like Sarvam AI and Krutrim are developing multilingual large language models tailored to...
Centerbase and Billables AI Give Midsize Law Firms Automated Time Capture and Actionable Practice Intelligence
Centerbase announced a native integration with NetDocuments' ndMAX AI, linking matter data directly to document intelligence for midsize law firms. The integration automatically extracts key contract details and writes them back into Centerbase, enabling real‑time reporting and workflow automation. It...
AGRO-AI Completes Technical Integration with WiseConn API
AGRO‑AI announced the completion of a technical integration with the WiseConn API, extending its irrigation intelligence engine to WiseConn‑controlled farms. The integration pulls farm, zone, telemetry, weather, and irrigation data, feeding it into AGRO‑AI’s recommendation, reporting, and execution‑assurance workflow. By...
Seven Ways AI Can Help During Your Annual Shareholders Meeting
The article outlines seven practical ways to deploy sandboxed AI during an annual shareholders meeting, from real‑time question handling to emergency response assistance. It suggests using AI to aggregate duplicate shareholder queries, gauge chat sentiment, and draft minutes with editable...
The 5-Prompt Sequence That Tells You If Your Ideas Hold Up
The post introduces a five‑prompt workflow—build, challenge, destroy, rebuild, decide—to rigorously test any business idea using AI tools like Claude or ChatGPT. The first two prompts steel‑man the concept and surface the three strongest objections, while the remaining steps dismantle,...
‘FOBO’ Is Spreading At Work — But The Timeline For AI Making You Obsolete May Surprise You
Workers are increasingly anxious about becoming obsolete, a condition dubbed FOBO, with 40 % citing AI‑driven job loss as a top fear, up from half that a year ago. MIT research shows AI can already complete 50‑75 % of text‑based tasks at...
I Built An Agentic ‘Law Firm’, Now What?
Antti Innanen launched Lavern, an AI‑driven "law firm" that runs 66 specialist agents on a single Mac Mini. The system mimics a traditional firm’s intake, decomposition, routing, debate, escalation and synthesis workflow, but all processing stays local, preserving client privacy....
Allies, Not Angels — Choosing to Embrace AI on Your Own Terms
Artificial intelligence is fundamentally a responsive tool built to extend human intent, not an autonomous rival. The essay argues that viewing AI as an "augment" rather than an "argument" reshapes how professionals collaborate with the technology. Fixed‑mindset resistance, rooted in...
Worth Reading: Shameless Guesses, Not Hallucinations
Scott Alexander argues that AI outputs are better described as "shameless guesses" or bullshit rather than hallucinations, because training rewards correct answers but does not penalize errors. He notes that saying "I don’t know" is discouraged for commercial reasons, turning...
How Can AI Augment the Architect for Scale and Productivity Improvements?
The article explores how artificial intelligence is transforming the architect’s function across business, enterprise, solution, and technical layers, delivering measurable productivity gains. It details AI‑driven benefits such as rapid content generation, richer stakeholder visualizations, and automated compliance mapping, while flagging...
The Intelligence Factory War
The Wall Street Journal published confidential financial documents from OpenAI and Anthropic, exposing stark strategic differences between the two AI firms. OpenAI is doubling down on massive scaling to capture monopoly rents from artificial general intelligence, while Anthropic is prioritizing...
AI with Human Feelings? Anthropic’s Claude Edges Closer
Anthropic’s new study reveals that Claude Sonnet 4.5 contains distinct clusters of artificial neurons that encode human‑like emotions such as happiness, sadness, joy, and fear. These "functional emotions" activate in response to specific cues and noticeably steer the model’s responses...
My Head Is Spinning… And This Is The Real Problem With AI
The author reflects on how generative AI, while powerful, creates a flood of possibilities that can overwhelm decision‑making. Instead of merely asking what tasks to do, AI prompts the question of what could be done, leading to endless options and...
Claude Code Leak: Researchers Find First Vulnerability
Anthropic unintentionally published a source map that revealed roughly 512,000 lines of Claude Code's TypeScript source. Researchers used the leak to uncover a critical flaw allowing command‑chain bypass of the tool's deny‑rule system after 50 subcommands. The vulnerability could let...
Is China's Domestic AI Stack Coalescing?
China’s AI ecosystem is rapidly coalescing as domestic chipmakers capture 41% of the AI accelerator server market, aiming for 50% by year‑end, while firms like DeepSeek and Zhipu rewrite model code to run on Huawei silicon. Nvidia disclosed a $4.5 billion...
Anthropic Number One AI in Ranking and Revenue – Making $30 Billion Per Year
Anthropic announced a major compute expansion with Google and Broadcom, securing multiple gigawatts of next‑generation TPU capacity slated for 2027. The company’s annualized revenue run‑rate surged to $30 billion in April 2026, up from $9 billion at the end of 2025. Over...
![[AINews] Gemma 4 Crosses 2 Million Downloads](/cdn-cgi/image/width=1200,quality=75,format=auto,fit=cover/https://substackcdn.com/image/fetch/$s_!DbYa!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F73b0838a-bd14-46a1-801c-b6a2046e5c1e_1130x1130.png)
[AINews] Gemma 4 Crosses 2 Million Downloads
Gemma 4 hit roughly 2 million downloads in its first week, outpacing the 1.4 million downloads Gemma 2 achieved since its June 2024 launch and approaching Gemma 3’s 6.7 million‑year total. The model quickly became the top‑trending model on Hugging Face, with developers running it on Apple Silicon devices such...
Large Language Models in Trading: Models and Market Dynamics
The newsletter highlights two emerging research streams that apply large language models (LLMs) to finance. First, researchers fine‑tune open‑source LLMs and combine them with retrieval‑augmented generation (RAG) to fuse structured price data with unstructured news, achieving higher predictive accuracy and...
The Acp-First Claude Code Bridge Openclaw Users Need Now
Anthropic’s move to API‑key authentication has split OpenClaw’s front‑end channels from Claude‑code’s repo‑level execution, clarifying billing and session persistence. OpenClaw now outlines three lanes—API‑key, revived Claude CLI reuse, and legacy token profiles—while recommending fresh builds use the API‑key path. The...
When Will Anthropic Surpass NVIDIA?
Anthropic reported an additional $10 billion in revenue last month, putting its annual run rate at roughly $10 billion—achieved in under four years, far faster than traditional SaaS firms. By contrast, NVIDIA generates $215 billion in annual revenue and trades at a 22‑times...
Microsoft Brings Power Apps Closer to M365 Copilot with Structured Data Access
Microsoft announced a public‑preview feature that links Power Apps with Microsoft 365 Copilot, allowing users to interact with model‑driven apps through natural language. The integration lets users view, edit, and create records directly from the Copilot chat interface, eliminating the need to...
