Mojo AI Brief

# Cup of Mojo -- Daily AI Brief -- Wednesday, April 22, 2026 Broadcast-ready source document. Read the COLD OPEN verbatim, cover every numbered beat in the RUNDOWN in tier-appropriate airtime, close with the MOJO TAKE verbatim. APPENDIX is reference only. **Articles analyzed:** 10 **Sources monitored:** 9 **Beats to cover:** 10 --- ## EDITORIAL GUIDANCE FOR PODCAST NARRATION Read these rules before narrating. They are non-negotiable. 1. **Read the COLD OPEN verbatim.** Do not rewrite or paraphrase it. It is the episode's signature opening. 2. **Cover every numbered beat.** Beats are numbered 1 through 8. Do not skip any beat. Do not consolidate. Each beat gets its own moment. 3. **Honor the tier structure.** `deep` beats get longer treatment with full context. `standard` beats are structured but concise. `rapid_fire` beats are short and punchy. Use roughly 2 minutes for the deep beat, 1 minute per standard beat, 20-30 seconds per rapid-fire beat. 4. **Cite sources by name** when presenting a claim. Say "OpenAI announced..." not "a company announced". 5. **Use only the plain-English text in each beat.** Do not pull technical jargon from the APPENDIX. The appendix is reference material for context, not script content. If a beat does not mention a term, do not introduce it. 6. **Only use numbers that appear in a beat's own text.** Do not import statistics from the appendix. Omit rather than fabricate. 7. **Reference earlier beats when topics connect.** Each beat has a `callbacks` field listing earlier beat numbers it relates to. When narrating, explicitly link back: "Remember that supply chain attack from Beat 1? This next one shows how the downstream risk compounds." Callbacks create cohesion and prevent the episode from feeling like a list. 8. **Introduce one skeptical angle per deep or standard beat.** Phrases like "one caveat", "critics will point out", or "this is not yet peer-reviewed" create credibility. Rapid-fire beats can skip this. 9. **Use the pronunciation guide for every named person or company.** Do not guess pronunciations. 10. **Close with the MOJO TAKE outro.** Read it as the host's editorial perspective, not as a summary. --- ## PRONUNCIATION GUIDE The following names appear in today's content. Use these phonetic pronunciations: - **Dario Amodei** — pronounced *DAR-ee-oh ah-moh-DAY* - **Jensen Huang** — pronounced *JEN-sen HWAHNG* - **Anthropic** — pronounced *an-THROP-ik* --- ## COLD OPEN -- Read This Verbatim Read the HOOK line first, pause for a beat, then the TEASE. Do not rewrite. Do not paraphrase. Do not add any preamble. > **Hook:** Anthropic and Amazon just shook hands on five gigawatts of new compute. That is not a data center, that is a small country's power grid. > **Tease:** OpenAI is quietly swapping HTTP for WebSockets to make agents feel less like dial-up. Semafor says token bills are creeping up on payroll. And Crunchbase has a reality check for anyone pitching this quarter. --- ## TODAY'S RUNDOWN Cover every beat in order. Do not skip. Tier labels tell you how much airtime each beat deserves. ### Beat ? [DEEP] — OpenAI swaps HTTP for WebSockets in the Responses API and Codex gets a lot snappier **Source:** OpenAI Blog | https://openai.com/index/speeding-up-agentic-workflows-with-websockets **Hook (open with this):** OpenAI just quietly rewired the Codex agent loop, and if you're orchestrating agents you want to pay attention. **Plain English:** OpenAI replaced the usual request-response plumbing in the Responses API with a persistent WebSocket connection. They also added connection-scoped caching, so the model remembers context across turns on the same socket instead of reloading it every call. Result: less overhead per call, faster first token, cheaper long agent runs. **Stakes:** Keep shipping agents over plain HTTP and you're paying the handshake tax on every tool call while competitors ship loops that feel twice as fast. **Twist:** The speedup didn't come from a smarter model, it came from not throwing away the connection between turns, which is networking 101 from 1999. **Takeaway:** Long-running agents live or die on connection reuse, not model IQ, so audit your transport layer before you swap models again. ### Beat ? [STANDARD] — Anthropic and Amazon double down with up to 5 gigawatts of fresh AWS compute for Claude **Source:** Anthropic Blog | https://www.anthropic.com/news/anthropic-amazon-compute **Callbacks:** references Beat 1. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** Anthropic just locked in up to 5 gigawatts of new AWS capacity, and that's not a vanity number, that's Claude's runway. **Plain English:** Anthropic and Amazon expanded their deal for up to five gigawatts of compute, most of it Trainium-based, to serve Claude at scale. For context, one gigawatt runs a small city. Five gigawatts means Anthropic is betting Claude demand keeps climbing and they need the silicon locked down now, not next year. **Stakes:** Ignore this and you'll be guessing why your Claude latency and pricing swing next quarter when everyone else is fighting for the same tokens. **Twist:** The headline isn't the model, it's the power grid: Anthropic is now competing for megawatts the way it used to compete for researchers. **Takeaway:** Capacity deals are the new model announcements, so watch who's buying gigawatts if you want to know whose API will actually be up in 2026. ### Beat ? [STANDARD] — OpenAI launches Workspace Agents in ChatGPT, turning every knowledge worker into a junior agent builder **Source:** OpenAI Blog | https://openai.com/academy/workspace-agents **Hook (open with this):** OpenAI just dropped Workspace Agents, a ChatGPT-native playbook for automating repeatable team workflows without calling a developer. **Plain English:** OpenAI is teaching business users to build agents directly inside ChatGPT, wire them to tools like Gmail and Drive, and scale them across a team. It's basically Zapier with a brain, sold to the person who used to file the ticket. No LangGraph, no Temporal, no engineer required for the first mile. **Stakes:** If you sell custom agent builds to SMBs, your buyer just got a free in-house option and will ask why your invoice has four zeros. **Twist:** The moat isn't the agent anymore, it's the messy glue work OpenAI's demo quietly skips: approvals, audit logs, and what happens when the agent hallucinates a refund. **Takeaway:** Sell the plumbing and the guardrails, not the agent itself, because ChatGPT now ships the agent for free. ### Beat ? [STANDARD] — Semafor says AI token bills are starting to rival payroll, and CFOs can't forecast them **Source:** Semafor | https://www.semafor.com/article/04/22/2026/ai-tokens-may-be-starting-to-rival-labor-costs **Callbacks:** references Beat 2, Beat 3. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** Semafor just dropped a line that should make every CFO spit out their coffee: token spend is creeping up on labor costs. **Plain English:** Companies are burning so much money on API calls to OpenAI, Anthropic, and Google that tokens are showing up as a real line item next to salaries. The problem is nobody knows how to forecast it. Usage spikes when one team ships an agent, and suddenly your quarterly budget is toast. **Stakes:** Ignore this and your finance team gets blindsided by a seven-figure OpenAI invoice nobody signed off on. **Twist:** The scary part isn't the size of the bill, it's that tokens behave like a variable labor cost you can't fire or furlough. **Takeaway:** Put a token budget on every agent before you ship it, or your CFO will put a budget on you. ### Beat ? [RAPID_FIRE] — Crunchbase and MGV's Marc Schröder say vertical AI is where the record venture money is actually landing **Source:** Crunchbase News (AI) | https://news.crunchbase.com/venture/building-successful-startup-vertical-ai-schroder-mgv/ **Callbacks:** references Beat 3. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** Marc Schröder at MGV just told Crunchbase the quiet part out loud: horizontal AI is a bloodbath, vertical AI is where checks are clearing. **Plain English:** Venture had a record quarter, but the money isn't chasing another general chatbot. It's chasing founders who pick one industry, elder care, yachting, dental, whatever, and bolt AI to a real workflow. Schröder says build for acquisition from day one. **Stakes:** Pitch a horizontal AI tool in 2025 and you're competing with OpenAI's free tier and a thousand wrappers for the same seed check. **Twist:** The record funding number hides the fact that generalist AI startups are getting passed over while boring vertical plays are oversubscribed. **Takeaway:** Pick a vertical, own the workflow, and the term sheet follows. ### Beat ? [RAPID_FIRE] — Apple Neural Engine finally learns to run Mixture-of-Experts models without choking **Source:** arXiv cs.LG | https://arxiv.org/abs/2604.18788 **Callbacks:** references Beat 2, Beat 4. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** Apple's Neural Engine can now handle MoE models, and that's a big deal for anyone eyeing a Mac Studio as their inference box. **Plain English:** Researchers cracked the three things that made MoE models a nightmare on Apple's NPU: unpredictable expert routing, weird operators like top-k, and shape constraints. Translation: sparse models like Mixtral and DeepSeek can finally run cheap on Apple silicon instead of burning GPU hours in the cloud. **Stakes:** Sleep on this and you'll keep paying cloud GPU prices while your competitor ships the same agent on a $4K Mac Studio sitting in a closet. **Twist:** The bottleneck was never the model size, it was the routing logic that NPUs were never designed for. **Takeaway:** Edge inference on Apple silicon just got real for the models that actually matter. ### Beat ? [RAPID_FIRE] — Senator Patty Murray wants CMS to scrap the WISeR AI prior-auth pilot delaying care in Washington hospitals **Source:** STAT News (AI) | https://www.statnews.com/2026/04/22/cms-wiser-program-delays-care-washington-state-hospitals-senator-says/?utm_campaign=rss **Callbacks:** references Beat 1. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** Senator Patty Murray is calling on CMS to kill the WISeR pilot after Washington hospitals say the AI prior-auth tool is stalling care for seniors. **Plain English:** CMS is testing AI to approve or deny Medicare procedures before they happen. Washington state hospitals report the bot is slowing things down and delaying care. Murray's office wants the whole program scrapped, not tweaked. **Stakes:** If you sell AI into healthcare payers, the political cover for aggressive automation just got a lot thinner. **Twist:** The complaint isn't that the model is wrong, it's that it's slow, which is the one thing AI was supposed to fix. **Takeaway:** Latency in a regulated workflow is a policy problem, not an engineering problem. ### Beat ? [RAPID_FIRE] — Hugging Face and NVIDIA ship a recipe for grounding Korean agents in real demographics using Nemotron synthetic personas **Source:** Hugging Face Blog | https://huggingface.co/blog/nvidia/build-korean-agents-with-nemotron-personas **Callbacks:** references Beat 5. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** Hugging Face and NVIDIA just dropped a playbook for building Korean agents that actually sound like Korean users, not a Seoul-themed Renaissance Faire. **Plain English:** NVIDIA's Nemotron Personas dataset lets you seed an agent with synthetic users whose age, region, and job mix match real Korean census data. You plug the personas into evals and fine-tuning so your agent stops defaulting to a 28-year-old Gangnam office worker for every query. It is persona grounding, not prompt hacking. **Stakes:** Ship a Korean agent trained on English assumptions and your retention in Busan will tell you exactly how badly you missed. **Twist:** The unlock isn't more Korean tokens, it's demographically weighted fake people telling you where your model is tone-deaf. **Takeaway:** If you're going vertical or going global, synthetic personas are now part of the eval stack, not a nice-to-have. ### Beat ? [RAPID_FIRE] — Anthropic drops Claude Opus 4.7 and Zvi Mowshowitz is already chewing through the model card **Source:** Zvi Mowshowitz | https://thezvi.substack.com/p/opus-47-part-1-the-model-card **Callbacks:** references Beat 2. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** Anthropic shipped Claude Opus 4.7 less than a week after Zvi finished writing up the last one, and he's back at the keyboard. **Plain English:** Zvi Mowshowitz is mid-series on Claude Mythos when Anthropic drops Opus 4.7 and its model card. Part 1 digs into the safety evals and capability jumps before anyone's even finished benchmarking the last version. **Stakes:** If your stack is pinned to an older Opus, you're one Tuesday away from a coworker asking why the new one is cheaper and smarter. **Twist:** Anthropic's release cadence is now faster than the independent analysts can publish on the previous model. **Takeaway:** Read the model card before you read the tweets, and read Zvi before you read the model card. ### Beat ? [RAPID_FIRE] — AutomationBench arrives to grade agents on CRM, inbox, calendar, and messaging coordination with policy rules attached **Source:** arXiv cs.AI | https://arxiv.org/abs/2604.18934 **Callbacks:** references Beat 3, Beat 4. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** AutomationBench just dropped on arXiv, and it finally tests the thing SMB agents actually do all day. **Plain English:** The benchmark scores agents on cross-app workflows that span a CRM, an inbox, a calendar, and a chat tool, all while following a written policy doc. It also grades whether the agent can discover the right APIs on its own, not just call ones you hand-wired. **Stakes:** If your agent demo only touches one app, you're grading on easy mode and your client's real workflow will expose it in week two. **Twist:** The hard part isn't the model, it's policy adherence across four systems without writing garbage into any of them. **Takeaway:** Run your agent against AutomationBench before your customer runs it against payroll. --- ## NOT WORTH YOUR TIME TODAY Do not cover on air. These are listed so the host can acknowledge if asked. - **ClawCoin: An Agentic AI-Native Cryptocurrency for Decentralized Agent Economies** -- A paper that staples 'agentic' to 'cryptocurrency' in the title is doing your filtering for you. Hard pass. - **HadAgent: Harness-Aware Decentralized Agentic AI Serving with Proof-of-Inference Blockchain Consensus** -- Four buzzwords, one acronym, zero shipping customers. Blockchain consensus for inference is a solution hunting for a problem. - **Quantum-inspired qubit-qutrit neural networks for real-time financial forecasting** -- 'Quantum-inspired' means 'not quantum.' If it actually beat a boring LSTM on real money, a hedge fund would own it, not arXiv. --- ## ACTION ITEMS FOR THIS WEEK (Joey only) These are internal action items. Not for on-air narration. - Audit Mojo's Responses API calls and rip out any lingering HTTP polling for WebSockets before Codex-style latency wins eat our lunch. Ping the platform team Monday and get a transport diff on my desk by Thursday. - Put a hard token budget on every Mojo agent shipping this quarter, then hand Finance a per-agent burn dashboard so our CFO stops reading Semafor and panicking about payroll-sized inference bills. - Run our top three customer-facing agents through AutomationBench this week on CRM, inbox, and calendar tasks, and publish the scores internally before sales quotes another demo number. --- ## MOJO TAKE -- Editorial Outro (Read Verbatim) Three-paragraph outro. Read each block verbatim, with natural pauses between them. > **Connect the dots:** Three threads braid today. OpenAI's WebSockets fix and Apple's MoE win say infrastructure beats IQ. Anthropic-AWS gigawatts and Semafor's token-bill panic say compute is the real P&L line. And Workspace Agents plus AutomationBench plus Schröder's vertical thesis all scream the same thing: the moat is the workflow, not the model. > **Watch next:** Watch whether Claude Opus 4.7 benchmarks hold up once Zvi finishes the model card, whether CMS actually blinks on WISeR after Murray's letter, and whether any CFO publicly prints a token budget next to payroll on the Q4 call. > **Sign-off:** That's your cup. Audit your transport, budget your tokens, own your vertical, and for the love of Sam Altman, read the model card. Joey out. --- ## APPENDIX -- VERBATIM SOURCE CONTENT Reference material. Do not read verbatim. Do not pull jargon from here into the spoken script. If the rundown beat does not mention a term, do not introduce it on the podcast. ### OpenAI swaps HTTP for WebSockets in the Responses API and Codex gets a lot snappier **Source:** OpenAI Blog **Link:** https://openai.com/index/speeding-up-agentic-workflows-with-websockets *RSS summary:* A deep dive into the Codex agent loop, showing how WebSockets and connection-scoped caching reduced API overhead and improved model latency. ### Anthropic and Amazon double down with up to 5 gigawatts of fresh AWS compute for Claude **Source:** Anthropic Blog **Link:** https://www.anthropic.com/news/anthropic-amazon-compute Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute We have signed a new agreement with Amazon that will deepen our existing partnership and secure up to 5 gigawatts (GW) of capacity for training and deploying Claude, including new Trainium2 capacity coming online in the first half of this year and nearly 1GW total of Trainium2 and Trainium3 capacity coming online by the end of 2026. We have worked closely with Amazon since 2023 and over 100,000 customers now run Claude on Amazon Bedrock. Together we launched Project Rainier, one of the largest compute clusters in the world, and we currently use over one million Trainium2 chips to train and serve Claude. Today’s agreement expands our collaboration in three ways. Infrastructure at scale. We are committing more than $100 billion over the next ten years to AWS technologies, securing up to 5GW of new capacity to train and run Claude. The commitment spans Graviton and Trainium2 through Trainium4 chips, with the option to purchase future generations of Amazon’s custom silicon as they become available. Significant Trainium2 capacity is coming online in Q2 and scaled Trainium3 capacity is expected to come online later this year. Anthropic will also use incremental capacity for Claude in Amazon Bedrock. The agreement includes expansion of inference in Asia and Europe to better serve Claude’s growing international customer base. We continue to choose AWS as our primary training and cloud provider for mission-critical workloads. “Our custom AI silicon offers high performance at significantly lower cost for customers, which is why it’s in such hot demand,” said Andy Jassy, CEO of Amazon. “Anthropic's commitment to run its large language models on AWS Trainium for the next decade reflects the progress we've made together on custom silicon, as we continue delivering the technology and infrastructure our customers need to build with generative AI.” Claude Platform on AWS. The full Claude Platform will be available directly within AWS. Same account, same controls, same billing, with more Claude Platform features and no additional credentials or contracts necessary. This gives organizations direct access to Claude while meeting their existing governance and compliance requirements. Claude remains the only frontier AI model available to customers on all three of the world's largest cloud platforms: AWS (Bedrock), Google Cloud (Vertex AI), and Microsoft Azure (Foundry). Claude Platform on AWS is coming soon. Reach out to your account team to request access. Continued investment. Amazon is investing $5 billion in Anthropic today, with up to an additional $20 billion in the future. This builds on the $8 billion Amazon has previously invested. “Our users tell us Claude is increasingly essential to how they work, and we need to build the infrastructure to keep pace with rapidly growing demand,” said Dario Amodei, CEO and co-founder of Anthropic. “Our collaboration with Amazon will allow us to ### OpenAI launches Workspace Agents in ChatGPT, turning every knowledge worker into a junior agent builder **Source:** OpenAI Blog **Link:** https://openai.com/academy/workspace-agents *RSS summary:* Learn how to build, use, and scale workspace agents in ChatGPT to automate repeatable workflows, connect tools, and streamline team operations. ### Apple Neural Engine finally learns to run Mixture-of-Experts models without choking **Source:** arXiv cs.LG **Link:** https://arxiv.org/abs/2604.18788 Computer Science > Machine Learning Title:Efficient Mixture-of-Experts LLM Inference with Apple Silicon NPUs View PDF HTML (experimental)Abstract:Apple Neural Engine (ANE) is a dedicated neural processing unit (NPU) present in every Apple Silicon chip. Mixture-of-Experts (MoE) LLMs improve inference efficiency via sparse activation but are challenging for NPUs in three ways: expert routing is unpredictable and introduces dynamic tensor shapes that conflict with the shape-specific constraints of NPUs; several irregular operators, e.g., top-k, scatter/gather, etc., are not NPU-friendly; and launching many small expert kernels incurs substantial dispatch and synchronization overhead. NPUs are designed to offload AI compute from CPU and GPU; our goal is to enable such offloading for MoE inference, particularly during prefill, where long-context workloads consume substantial system resources. This paper presents NPUMoE, a runtime inference engine that accelerates MoE execution on Apple Silicon by offloading dense, static computation to NPU, while preserving a CPU/GPU fallback path for dynamic operations. NPUMoE uses offline calibration to estimate expert capacity and popularity that drives three key techniques: (1) Static tiers for expert capacity to address dynamic expert routing (2) Grouped expert execution to mitigate NPU concurrency limits (3) Load-aware expert compute graph residency to reduce CPU-NPU synchronization overhead. Experiments on Apple M-series devices using three representative MoE LLMs and four long-context workloads show that NPUMoE consistently outperforms baselines, reducing latency by 1.32x-5.55x, improving energy efficiency by 1.81x-7.37x, and reducing CPU-cycle usage by 1.78x-5.54x through effective NPU offloading. ### AutomationBench arrives to grade agents on CRM, inbox, calendar, and messaging coordination with policy rules attached **Source:** arXiv cs.AI **Link:** https://arxiv.org/abs/2604.18934 Computer Science > Artificial Intelligence Title:AutomationBench View PDF HTML (experimental)Abstract:Existing AI benchmarks for software automation rarely combine cross-application coordination, autonomous API discovery, and policy adherence. Real business workflows demand all three: a single task may span a CRM, inbox, calendar, and messaging platform - requiring the agent to find the right endpoints, follow a policy document, and write correct data to each system. To address this gap, we introduce AutomationBench, a benchmark for evaluating AI agents on cross-application workflow orchestration via REST APIs. Drawing on real workflow patterns from Zapier's platform, tasks span Sales, Marketing, Operations, Support, Finance, and HR domains. Agents must discover relevant endpoints themselves, follow layered business rules, and navigate environments with irrelevant and sometimes misleading records. Grading is programmatic and end-state only: whether the correct data ended up in the right systems. Even the best frontier models currently score below 10%. AutomationBench provides a challenging, realistic measure of where current models stand relative to the agentic capabilities businesses actually need. ### Semafor says AI token bills are starting to rival payroll, and CFOs can't forecast them **Source:** Semafor **Link:** https://www.semafor.com/article/04/22/2026/ai-tokens-may-be-starting-to-rival-labor-costs The cost of tokens is now competing with the cost of headcount. “You have to raise a lot more money on a per-headcount basis,” Henry Ward, who runs Silicon Valley financial-software provider Carta, said at Semafor World Economy in Washington, DC. “I don’t see any end to that in sight.” Tech executives like Nvidia’s Jensen Huang have suggested that the more companies spend on AI, the more money they will make. But the growing cost of tokens is a line item that CFOs are having trouble planning for. Several executives in Washington, DC, last week said they’re grappling with that uncertainty. “The unit costs are going down, but the aggregate costs are going up, and companies don’t like when something is unpredictable on cost,” said Charles Phillips, co-founder of private equity firm Recognize. Kunal Kapoor, CEO of Morningstar, added that subscription models “have endured for a long time because there’s value in certainty.” ICONIQ Capital founding partner Divesh Makan, whose investment firm recently led one of Anthropic’s funding rounds, called free AI models “the gateway drug” to paid versions, adding that his employees keep asking for more tokens. “The conversation we’re having is, ‘What is the amount we should be spending, and how do we measure ROI?’” he said. “Are you just buying holidays and checking the weather in Tokyo, or are you doing something productive with these tokens?” ### Hugging Face and NVIDIA ship a recipe for grounding Korean agents in real demographics using Nemotron synthetic personas **Source:** Hugging Face Blog **Link:** https://huggingface.co/blog/nvidia/build-korean-agents-with-nemotron-personas Nemotron-Personas-Korea fixes this. The dataset provides 6 million fully synthetic personas grounded in official statistics and seed data from the Korean Statistical Information Service (KOSIS), the Supreme Court of Korea, the National Health Insurance Service, and the Korea Rural Economic Institute. NAVER Cloud contributed seed data and domain expertise during design. Every persona is demographically accurate but contains zero personally identifiable information (PII). It’s designed with Korea's Personal Information Protection Act (PIPA) in mind. South Korea is also one of the few countries to publish an official Synthetic Data Generation guide, establishing governance for grounding models with synthetic versions of sensitive data. This dataset follows that approach. In this tutorial, we'll turn a synthetic persona into a deployed Korean agent — from filtering the dataset to inference — in about 20 minutes using hosted APIs. Nemotron-Personas-Korea was generated using NeMo Data Designer, NVIDIA's open-source compound AI system for synthetic data. The pipeline pairs a Probabilistic Graphical Model (Apache-2.0) for statistical grounding with Gemma-4-31B for Korean-language narrative generation. Population data comes from KOSIS (2020–2026 releases); name distributions come from the Supreme Court of Korea. Nemotron-Personas-Korea is the latest addition to the Nemotron-Personas Collection, which also covers the USA, Japan, India, Singapore (with AI Singapore), Brazil (with WideLabs), and France (with Pleias). If you're building a multilingual agent that serves Korean users alongside other markets, you can blend personas across countries in the same pipeline. Most agents today are identity-blind. They follow instructions without any grounding in who they're serving. For example, an agent that books a Korean hospital appointment using US scheduling conventions, or addresses a 60-year-old patient in 반말 (“banmal,” informal language), doesn't just feel wrong. It fails. Nemotron-Personas-Korea changes this by giving your agent a Korean operating context. Load a persona into the system prompt and the agent inherits that persona's region, occupation, communication norms, and domain expertise. This works across any agent framework. Deploy with NemoClaw (NVIDIA's open-source reference stack for always-on agents running in NVIDIA OpenShell sandboxes, on anything from RTX PCs to DGX Spark), serve through NVIDIA NIM for production inference, or call the NVIDIA API directly. The persona layer is framework-agnostic, acting as a well-structured system prompt grounded in real Korean demographics. 🔗 Resources - Nemotron-Personas-Korea for seeding training - NeMo Data Designer for synthesizing domain-specific data - NVIDIA NemoClaw for deploying always-on agents - NVIDIA Developer Discord for community support Load the dataset and explore what's available. Each record contains structured demographic fields alongside rich, natural-language persona narratives. from datas ### Anthropic drops Claude Opus 4.7 and Zvi Mowshowitz is already chewing through the model card **Source:** Zvi Mowshowitz **Link:** https://thezvi.substack.com/p/opus-47-part-1-the-model-card Opus 4.7 Part 1: The Model Card Less than a week after completing coverage of Claude Mythos, here we are again as Anthropic gives us Claude Opus 4.7. So here we are, with another 232 pages of light reading. This post covers the first six sections of the Model Card. It excludes section seven, model welfare, because there are concerns this time around that need to be expanded into their own post. The reason model welfare and related topics get their own post this time around is that some things clearly went seriously wrong on that front, in ways they haven’t gone wrong in previous Claude models. Tomorrow’s post is in large part an investigation of that, as best I can from this position, including various hypotheses for what happened. This post also excludes section eight, capabilities, which will be included in the capabilities and reactions post as per usual. Consider this the calm before the storm. Since I likely won’t get to capabilities until Wednesday, for those experiencing first contact with Opus 4.7, a few quick tips: Turning off ‘adaptive thinking’ means no thinking, period. Terrible UI. So make sure to keep this on. If you need it to definitely think, you can do that via Claude Code, which can do non-code things too. On Claude Code Opus 4.7 now defaults to xhigh thinking, which will eat a lot of tokens. If you’re at risk of running out you might not want that. You probably do want to have it in auto mode. You need, more so than usual, to ‘treat the model well’ if you want good results. Treat it like a coworker, and do not bark orders or berate it.Different people get more different experiences than with prior models. Your system instructions may no longer be helping. Consider changing them. There were some bugs that have been fixed. If you encountered issues in the first day or two, consider trying again. All right, let’s go. Here We Go Again: Executive Summary This is my summary of their summary, plus points I would have put in a summary. Mythos exists, so if one includes it then I presume Anthropic are right that Claude Opus 4.7 is not advancing the capability frontier, so if Mythos doesn’t set off the RSP triggers then one can assume Opus 4.7 shouldn’t either. Capabilities are ahead of 4.6, well behind Mythos. Cyber for Opus 4.7 is similar to Opus 4.6. This is no Mythos. Mundane safety is solid and similar to Opus 4.6. Opus 4.7 is more robust to prompt injections and during computer use. Model welfare self-reports and internal emotion representations are positive. I don’t think they did a good job summarizing their findings here, and I will be covering model welfare and related issues in their own post this time around. Introduction (1) Claude Opus 4.7 was trained on the standard Anthropic training things, in the standard ways, and evaluated in the standard ways. The release decision was that Opus 4.7 was not substantially different than Opus 4.6 on any of the key risk dimensions. I notice they still haven’t updated to include cyber in ### Crunchbase and MGV's Marc Schröder say vertical AI is where the record venture money is actually landing **Source:** Crunchbase News (AI) **Link:** https://news.crunchbase.com/venture/building-successful-startup-vertical-ai-schroder-mgv/ Crunchbase just reported that $300 billion flowed into startups in Q1 2026, the biggest quarter in venture history. The eye-popping subtext? Four companies absorbed $188 billion of that, or 65%. If you’re a seed-stage founder reading those numbers, it’s easy to feel like the market is passing you by. Look closer, and the story changes completely. Early-stage funding was up 41% year over year. AI/ML deal count rose to 6,678 in 2025, up from roughly 5,600 the year before. More companies are getting funded at the early stage, not fewer. The concentration at the top? That’s an infrastructure play. The application layer looks entirely different. Build vertical, not horizontal The real signal is in the shift from horizontal to vertical. Redpoint’s 2026 Market Update shows horizontal SaaS down 35% over the past 12 months while vertical SaaS is essentially flat (up 3%). That divergence matters for founders deciding what to build. Horizontal software (project management, general productivity, collaboration) is commoditizing fast as AI agents handle coordination natively. But vertical software? That’s where proprietary data shines and industry-specific compliance workflows matter. AI makes the first category less valuable and the second category more valuable. If you’re starting a company right now, the data says: Pick an industry, not a feature. Claims processing in insurance, scheduling in healthcare, compliance in financial services, job costing in construction. These are workflows where software penetration has been shallow for decades because the problems were too specific for horizontal tools. AI changes that math. Build for the $6T, not the $500B The addressable market for software is expanding, not contracting. In Redpoint’s CIO survey, 58% cite AI as the top driver of increased software spend. As agents move from copilot features into autonomous workflow execution, the addressable market grows from roughly $0.5 trillion in current U.S. enterprise software spend toward $6 trillion or more, because AI starts capturing portions of the knowledge-worker payroll that software never could. This is classic Jevons’ Paradox: When a resource gets dramatically cheaper to produce, consumption goes up. AI makes software dramatically cheaper to build, deploy and maintain. Suddenly, job costing for midsize contractors pencils out. Inventory optimization for independent pharmacies becomes economically viable. The cottage industries that enterprise software ignored for decades? They’re all in play now. Build for acquirability, not just IPO optionality But let’s talk about exits, because that’s where the rubber meets the road. The IPO market remains largely closed. In 2025, roughly 2,300 VC-backed startups were acquired compared to just 65 IPOs, per Crunchbase data. LPs have seen nearly $200 billion in cumulative negative net cash flows since 2022. The pressure to return capital through M&A is real and growing. Smart founders are building for this reality from day o ### Senator Patty Murray wants CMS to scrap the WISeR AI prior-auth pilot delaying care in Washington hospitals **Source:** STAT News (AI) **Link:** https://www.statnews.com/2026/04/22/cms-wiser-program-delays-care-washington-state-hospitals-senator-says/?utm_campaign=rss Washington state hospitals say their Medicare patients are waiting two to four times longer in some cases for procedures that are now subject to prior authorization under a new Medicare program. The report from Sen. Maria Cantwell (D-Wash.) is among the first to document alleged patient harm stemming from the Centers for Medicare and Medicaid Services’ new Wasteful and Inappropriate Service Reduction, or WISeR, Model. Cantwell is one of several Democratic members of Congress who have been urging CMS to scrap the program, which launched Jan. 1. Cantwell aired her concerns about WISeR to Health and Human Services Secretary Robert F. Kennedy Jr. at a Senate Committee on Finance hearing Wednesday. She said CMS is using artificial intelligence as a “denial device” and that patients are waiting weeks to get sign off for services that previously didn’t require approval. This article is exclusive to STAT+ subscribers Unlock this article — plus daily intelligence on Capitol Hill and the life sciences industry — by subscribing to STAT+. Already have an account? Log in