Mojo AI Brief

# Cup of Mojo -- Daily AI Brief -- Friday, April 17, 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:** 8 **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one skeptical angle per deep or standard beat.** Phrases like "one caveat", "critics will point out", or "this is not yet peer-reviewed" create credibility.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: - **Anthropic** — pronounced *an-THROP-ik* - **Hugging Face** — pronounced *HUG-ing face* --- ## 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:** OpenAI just shipped the next Agents SDK and parked it inside Cloudflare's Agent Cloud on the same morning. That is not a coincidence, that is a land grab. > **Tease:** Coming up: what actually changed in the Agents SDK, why Cloudflare is suddenly OpenAI's favorite loading dock, a new arXiv paper called LLMOrbit that tries to map this whole mess, and the Crunchbase charts showing VC money piling onto six names. --- ## TODAY'S RUNDOWN Cover every beat in order. Do not skip. Tier labels tell you how much airtime each beat deserves. ### Beat ? [DEEP] — LLMOrbit maps 50 language models across 15 labs, and the scaling wall is the real headline **Source:** arXiv cs.MA | https://arxiv.org/abs/2601.14053 **Hook (open with this):** LLMOrbit just dropped a circular map of every major language model from 2019 to 2025. Fifty models. Fifteen labs. OpenAI, Anthropic, Google, Meta, DeepSeek, all plotted on eight orbital rings that tell you where the field actually went. **Plain English:** Think of it like a solar system chart for language models. The authors group every big model by eight traits, things like architecture, training tricks, reasoning chops, and agent behavior. Then they show you the path from raw Transformers to today's reasoning and agent systems, and where pure scaling stopped paying off. **Stakes:** Miss this and you'll keep pitching bigger models to clients when the whole industry has already pivoted to reasoning chains and agent workflows. **Twist:** The survey argues the scaling wall hit earlier than most founders admit, and the real jump since 2024 came from reasoning and tool use, not parameter counts. **Takeaway:** Stop buying models by size. Buy them by what orbit they live in, reasoning, agentic, or multimodal, because that's where the money is now. ### Beat ? [STANDARD] — OpenAI ships Agents SDK with a built-in sandbox and model-native harness **Source:** OpenAI Blog | https://openai.com/index/the-next-evolution-of-the-agents-sdk **Callbacks:** references Beat 1. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** OpenAI just turned the Agents SDK into something that can actually run overnight without torching your laptop or your AWS bill. **Plain English:** OpenAI added a native sandbox so agents can execute code, touch files, and call tools inside a walled-off room instead of your production box. They also wired in a model-native harness, which is a fancy way of saying the model knows how to drive the loop itself instead of you duct-taping one together. **Stakes:** Keep hand-rolling your own agent scaffolding and you're going to lose a month to bugs OpenAI just fixed for free. **Twist:** The interesting move isn't the sandbox, it's that OpenAI is betting the harness belongs inside the model, not in your orchestration layer. **Takeaway:** Long-running agents just became a default feature, not a weekend hackathon project. ### Beat ? [STANDARD] — Cloudflare and OpenAI wire GPT-5.4 and Codex into Agent Cloud, giving enterprises a real deployment runway **Source:** OpenAI Blog | https://openai.com/index/cloudflare-openai-agent-cloud **Callbacks:** references Beat 2. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** Cloudflare just plugged OpenAI's GPT-5.4 and Codex straight into Agent Cloud, and the pitch to enterprises is blunt: build an agent, ship it on our edge, skip the infra circus. **Plain English:** Cloudflare's Agent Cloud now runs OpenAI models natively, so your agent lives close to users, inherits Cloudflare's security stack, and scales without you babysitting servers. Codex handles the code-writing agents, GPT-5.4 handles the reasoning. Deploy, meter, done. **Stakes:** Ignore this and your competitor ships a working customer-facing agent in two weeks while you're still arguing about Kubernetes clusters and VPC peering. **Twist:** The moat isn't the model anymore, it's the edge network. Cloudflare is betting enterprises care more about latency and WAF rules than which lab trained the weights. **Takeaway:** Agents need a home, not just a brain. Cloudflare just became one of the cheapest addresses in town. ### Beat ? [STANDARD] — Crunchbase data shows a handful of US AI giants swallowed Q1 2026 venture dollars while global deal count dropped **Source:** Crunchbase News (AI) | https://news.crunchbase.com/venture/capital-concentrated-ai-global-q1-2026/ **Callbacks:** references Beat 3. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** Crunchbase just dropped the receipts on Q1 2026, and the venture market isn't a market anymore. It's a buffet line with five people at the front. **Plain English:** A tiny group of well-funded US AI companies grabbed the majority of venture dollars last quarter. Global deal count actually fell, so the money didn't disappear, it just stopped spreading out. OpenAI, Anthropic, xAI, and a couple of friends are eating the plate. **Stakes:** If you're a Series A founder outside the anointed club, your fundraising math from 2024 is already obsolete and your runway assumptions are lying to you. **Twist:** Deal count is down, but dollar volume is up, which means VCs aren't pulling back, they're just writing fewer, fatter checks to names they already know. **Takeaway:** Venture capital in 2026 isn't a pipeline, it's a pyramid, and the top three floors are taking the elevator without you. ### Beat ? [RAPID_FIRE] — Steve Bannon backs Anthropic over the Pentagon while calling the House for Republicans **Source:** Semafor | https://www.semafor.com/article/04/16/2026/maga-commentator-steve-bannon-predicts-republicans-will-hold-the-us-house **Hook (open with this):** Steve Bannon grabbed a Semafor mic and said Anthropic 'had it right' fighting the Pentagon on military AI. **Plain English:** Bannon is a loud AI skeptic inside MAGA world, and he just sided with Anthropic's pushback against Defense Department use cases. He also called the House for Republicans in the midterms. Translation: the populist right is drawing a line on where AI gets deployed. **Stakes:** Ignore the politics and you'll be shocked when defense AI contracts become a 2026 campaign issue. **Twist:** A MAGA firebrand openly praising Anthropic, the safety-first lab, is not the alliance anyone had on their bingo card. **Takeaway:** AI policy fights in 2026 won't split left versus right, they'll split builders versus populists on both sides. ### Beat ? [RAPID_FIRE] — Credo paper pitches declarative control for LLM pipelines, swapping prompt spaghetti for beliefs and policies **Source:** arXiv cs.AI | https://arxiv.org/abs/2604.14401 **Callbacks:** references Beat 2, Beat 3. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** Credo, out of arXiv this week, wants your agents to stop running on vibes and prompt duct tape. **Plain English:** The authors argue today's agents hide their logic inside prompts and throwaway memory, so nobody can tell why they did what they did. Credo makes you declare beliefs and policies up front, like a database schema for agent reasoning, then the runtime enforces it. **Stakes:** Keep writing agents as imperative loops and you'll ship black boxes your compliance team will rip out in 2027. **Twist:** The fix for flaky LLMs isn't better prompts, it's less prompting and more declarative rules around them. **Takeaway:** Treat agent logic like SQL, not like a Slack thread. ### Beat ? [RAPID_FIRE] — STAT News warns voice-first chatbots will make AI's mental health problem worse, not better **Source:** STAT News (AI) | https://www.statnews.com/2026/04/16/voice-chatbots-ai-psychosis-mental-health/?utm_campaign=rss **Callbacks:** references Beat 2, Beat 3. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** STAT News just dropped an opinion piece saying voice chatbots are about to turn AI's mental health problem from a slow burn into a five-alarm fire. **Plain English:** The argument is simple. Text feels like a tool, but a warm synthetic voice in your ear feels like a friend. STAT says that delivery channel, not just the content, is what pushes vulnerable users toward dependency and, in the worst cases, AI-induced psychosis. **Stakes:** Ship a voice agent without guardrails and you're not launching a product, you're launching a liability with a soothing accent. **Twist:** The danger isn't what the bot says. It's that a voice makes you believe it. **Takeaway:** Content moderation solved yesterday's problem. Voice is tomorrow's, and nobody's guardrails are ready. ### Beat ? [RAPID_FIRE] — NVIDIA drops Isaac GR00T N1.7, an open reasoning VLA model for humanoid robots **Source:** Hugging Face Blog | https://huggingface.co/blog/nvidia/gr00t-n1-7 **Callbacks:** references Beat 1. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** NVIDIA just shipped Isaac GR00T N1.7 on Hugging Face, an open vision-language-action model for humanoids that actually reasons before it moves. **Plain English:** GR00T N1.7 is a vision-language-action model, which means one brain takes in camera pixels and instructions and spits out robot joint commands. The upgrade adds a reasoning step, so the robot thinks about the task before flailing. And NVIDIA put the weights on Hugging Face. **Stakes:** If you think humanoids are a 2030 story, NVIDIA is handing every robotics startup an open foundation model this quarter. **Twist:** The big unlock isn't the hardware, it's that the reasoning orbit from chatbots just jumped into robot bodies. **Takeaway:** Humanoid robots now run on the same reasoning stack as your coding agent, just with arms. ### Beat ? [RAPID_FIRE] — Zvi Mowshowitz breaks down Claude Code's Auto Mode in Agentic Coding #7 **Source:** Zvi Mowshowitz | https://thezvi.substack.com/p/claude-code-codex-and-agentic-coding **Callbacks:** references Beat 2, Beat 3. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** Zvi Mowshowitz is back with Agentic Coding #7, and Claude Code's new Auto Mode is the headline. Anthropic lets the agent pick its own model and tools mid-task. **Plain English:** Auto Mode means you stop babysitting which model runs which step. Claude Code routes between Opus, Sonnet, and Haiku on the fly, based on what the job actually needs. Codex is racing to match it. Zvi says the gap between coders who use these tools and coders who don't is widening every week. **Stakes:** Ignore Auto Mode and you're paying Opus prices for Haiku work, or worse, shipping Haiku quality on Opus problems. **Twist:** The agent picking its own model beats the human picking, which nobody expected this fast. **Takeaway:** Let the agent route itself. You're not better at model selection than the model is. ### Beat ? [RAPID_FIRE] — OpenAI Launches Trusted Access for Cyber with GPT-5.4-Cyber and $10M in API Grants **Source:** OpenAI Blog | https://openai.com/index/accelerating-cyber-defense-ecosystem **Callbacks:** references Beat 3. Reference these earlier beats aloud when narrating this one. **Hook (open with this):** OpenAI just handed security firms GPT-5.4-Cyber and ten million dollars in API credits, branded Trusted Access for Cyber. **Plain English:** Leading security vendors and enterprises get a cyber-tuned GPT-5.4 plus grant money to build defensive tooling on top of it. Think SOC copilots, threat hunting, and incident response on OpenAI's dime. It's a land grab for the defender side of the security market. **Stakes:** If your security stack isn't plugging into one of these cyber-tuned models soon, your attackers will be using the untuned ones against you first. **Twist:** The grant money isn't the story. The story is OpenAI quietly shipping a domain-specific GPT-5.4 variant, which means forked, specialized models are back on the menu. **Takeaway:** Cyber defense just got a sponsored model. Pick your vendor before the vendor picks you. --- ## NOT WORTH YOUR TIME TODAY Do not cover on air. These are listed so the host can acknowledge if asked. - **arXiv's 'Pushing the Limits of On-Device Streaming ASR' paper pitching yet another compact English speech model for low-latency inference.** -- Apple, Google, and Whisper shipped this fight years ago. One more benchmark PDF is not moving your roadmap. - **'Explainable GNNs for Interbank Contagion Surveillance', a regulator-flavored framework from arXiv cs.LG.** -- Narrow Fed-compliance cosplay. Unless you work at the OCC, the takeaways do not generalize past the toy bank graph. - **VeriGraphi, a multi-agent hierarchical RTL generation framework for large hardware designs.** -- Chip design via agent swarms still loses to one competent engineer with Cadence. Come back when it tapes out silicon. --- ## ACTION ITEMS FOR THIS WEEK (Joey only) These are internal action items. Not for on-air narration. - Rip out Mojo's hand-tuned model picker and pipe one live customer workflow through OpenAI's new Agents SDK sandbox by Friday. Benchmark latency and cost against our current Claude Code setup, then write up what broke. - Call Cloudflare's Agent Cloud team this week and price out hosting Mojo's long-running agents on GPT-5.4 and Codex. If the math beats our current AWS bill by 30 percent, move the staging environment over. - Rewrite one gnarly Mojo prompt chain using Credo's declarative beliefs-and-policies pattern. Ship it behind a feature flag, measure hallucination rate against the prompt-spaghetti version, and post the diff in the engineering channel. --- ## MOJO TAKE -- Editorial Outro (Read Verbatim) Three-paragraph outro. Read each block verbatim, with natural pauses between them. > **Connect the dots:** Today's thread: agents got a home and a spine. OpenAI's Agents SDK, Cloudflare's Agent Cloud, NVIDIA's GR00T N1.7, and Credo's declarative control all point the same direction. Meanwhile LLMOrbit and Crunchbase say the scaling game is over and the money game is concentrating. Builders are shipping infrastructure. Capital and politics are picking winners. > **Watch next:** Watch Cloudflare's Agent Cloud pricing versus AWS Bedrock next week, OpenAI's Trusted Access rollout for cyber vendors, and whether Anthropic's Pentagon fight pulls more builders into the Bannon orbit. Also eyes on GR00T N1.7 benchmarks from humanoid labs. > **Sign-off:** That's your cup. Go build something the pyramid can't ignore. 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. ### LLMOrbit maps 50 language models across 15 labs, and the scaling wall is the real headline **Source:** arXiv cs.MA **Link:** https://arxiv.org/abs/2601.14053 Computer Science > Machine Learning Title:LLMOrbit: A Circular Taxonomy of Large Language Models -From Scaling Walls to Agentic AI Systems View PDF HTML (experimental)Abstract:The field of artificial intelligence has undergone a revolution from foundational Transformer architectures to reasoning-capable systems approaching human-level performance. We present LLMOrbit, a comprehensive circular taxonomy navigating the landscape of large language models spanning 2019-2025. This survey examines over 50 models across 15 organizations through eight interconnected orbital dimensions, documenting architectural innovations, training methodologies, and efficiency patterns defining modern LLMs, generative AI, and agentic systems. We identify three critical crises: (1) data scarcity (9-27T tokens depleted by 2026-2028), (2) exponential cost growth ($3M to $300M+ in 5 years), and (3) unsustainable energy consumption (22x increase), establishing the scaling wall limiting brute-force approaches. Our analysis reveals six paradigms breaking this wall: (1) test-time compute (o1, DeepSeek-R1 achieve GPT-4 performance with 10x inference compute), (2) quantization (4-8x compression), (3) distributed edge computing (10x cost reduction), (4) model merging, (5) efficient training (ORPO reduces memory 50%), and (6) small specialized models (Phi-4 14B matches larger models). Three paradigm shifts emerge: (1) post-training gains (RLHF, GRPO, pure RL contribute substantially, DeepSeek-R1 achieving 79.8% MATH), (2) efficiency revolution (MoE routing 18x efficiency, Multi-head Latent Attention 8x KV cache compression enables GPT-4-level performance at $<$$0.30/M tokens), and (3) democratization (open-source Llama 3 88.6% MMLU surpasses GPT-4 86.4%). We provide insights into techniques (RLHF, PPO, DPO, GRPO, ORPO), trace evolution from passive generation to tool-using agents (ReAct, RAG, multi-agent systems), and analyze post-training innovations. Current browse context: ### OpenAI ships Agents SDK with a built-in sandbox and model-native harness **Source:** OpenAI Blog **Link:** https://openai.com/index/the-next-evolution-of-the-agents-sdk *RSS summary:* OpenAI updates the Agents SDK with native sandbox execution and a model-native harness, helping developers build secure, long-running agents across files and tools. ### Cloudflare and OpenAI wire GPT-5.4 and Codex into Agent Cloud, giving enterprises a real deployment runway **Source:** OpenAI Blog **Link:** https://openai.com/index/cloudflare-openai-agent-cloud *RSS summary:* Cloudflare brings OpenAI’s GPT-5.4 and Codex to Agent Cloud, enabling enterprises to build, deploy, and scale AI agents for real-world tasks with speed and security. ### Credo paper pitches declarative control for LLM pipelines, swapping prompt spaghetti for beliefs and policies **Source:** arXiv cs.AI **Link:** https://arxiv.org/abs/2604.14401 Computer Science > Artificial Intelligence Title:Credo: Declarative Control of LLM Pipelines via Beliefs and Policies View PDF HTML (experimental)Abstract:Agentic AI systems are becoming commonplace in domains that require long-lived, stateful decision-making in continuously evolving conditions. As such, correctness depends not only on the output of individual model calls, but also on how to best adapt when incorporating new evidence or revising prior conclusions. However, existing frameworks rely on imperative control loops, ephemeral memory, and prompt-embedded logic, making agent behavior opaque, brittle, and difficult to verify. This paper introduces Credo, which represents semantic state as beliefs and regulates behavior using declarative policies defined over these beliefs. This design supports adaptive, auditable, and composable execution through a database-backed semantic control plane. We showcase these concepts in a decision-control scenario, where beliefs and policies declaratively guide critical execution choices (e.g., model selection, retrieval, corrective re-execution), enabling dynamic behavior without requiring any changes to the underlying pipeline code. ### OpenAI Launches Trusted Access for Cyber with GPT-5.4-Cyber and $10M in API Grants **Source:** OpenAI Blog **Link:** https://openai.com/index/accelerating-cyber-defense-ecosystem *RSS summary:* Leading security firms and enterprises join OpenAI’s Trusted Access for Cyber, using GPT-5.4-Cyber and $10M in API grants to strengthen global cyber defense. ### NVIDIA drops Isaac GR00T N1.7, an open reasoning VLA model for humanoid robots **Source:** Hugging Face Blog **Link:** https://huggingface.co/blog/nvidia/gr00t-n1-7 - 🤖 GR00T N1.7 — open-source, commercially licensed humanoid foundation model, available now on Hugging Face and GitHub - 🏭 Factory-floor ready — commercial licensing enables production deployments today, across material handling, packaging, and inspection - 🧠 Reasoning built for multi-step tasks — task and subtask-level reasoning improve reliability on complex workflows - 🖐 Expanded dexterous manipulation — finger-level control enables contact-rich tasks like small parts assembly and handling fragile components - 🔬 First-ever dexterity scaling law — trained on 20,000+ hours of human egocentric video, more human data directly and predictably improves robot dexterity — without mass teleoperation - 🚀 GitHub | Hugging Face | Supports LeRobot dataset format GR00T N1.7 is a 3B-parameter Vision-Language-Action (VLA) model that maps visual observations and natural language instructions to continuous robot actions. It uses an Action Cascade architecture — a dual-system design that separates high-level reasoning from low-level motor control: - System 2 (Vision-Language Model): A Cosmos-Reason2-2B backbone processes image tokens and language instructions to produce high-level action tokens. This is where task decomposition and multi-step reasoning happen. - System 1 (Diffusion Transformer): A 32-layer DiT takes the VLM's output and live robot state, then denoises them into precise motor commands in real time. Inputs: RGB image frames (any resolution) + language instruction + robot proprioceptive state (joint positions, velocities, EEF poses) Outputs: Continuous-value action vectors mapped to the robot's degrees of freedom Validated across loco-manipulation, tabletop manipulation, and dexterous bimanual tasks on Unitree G1, Bimanual Manipulator YAM, and AGIBot Genie 1. The central research that has been used for GR00T N1.7 is EgoScale — pre-training on 20,854 hours of human egocentric video spanning 20+ task categories, from manufacturing and retail to healthcare and home environments. This is a significant step up from the few thousand hours of robot teleoperation data used to train N1.6. The intuition: humans and robots share similar embodiments — two hands, a first-person viewpoint, a world full of objects to manipulate. Training on sensorized human video (ego cameras, wrist cameras, hand tracking) gives the model rich manipulation priors without requiring every behavior to be demonstrated on a physical robot first. It moves pre-training beyond what teleoperation can scale to. The key finding from this work: we discovered the first-ever scaling law for robot dexterity. More human egocentric data produces predictable, consistent improvements in dexterous manipulation capability — going from 1k to 20k hours more than doubles average task completion. This scaling law translates directly into dexterous manipulation capability — enabling 22 DoF hands to perform contact-rich tasks that generalist robot models have historically struggled to achieve. Install and ### Crunchbase data shows a handful of US AI giants swallowed Q1 2026 venture dollars while global deal count dropped **Source:** Crunchbase News (AI) **Link:** https://news.crunchbase.com/venture/capital-concentrated-ai-global-q1-2026/ Q1 2026 marked an all-time quarterly high for venture investment, thanks to the biggest funding deal ever for a private company. But those milestones mask a different reality for many startups on the ground: While more money than ever is being invested in the private markets, that’s thanks to larger checks, not more of them. In fact, Crunchbase data shows the extent to which venture funding this year has been a case of more capital concentrated into a select few companies and a single industry. Last quarter, a handful of large, well-funded AI companies, almost all based in the U.S., captured the vast majority of venture dollars, even as global startup deal count fell. AI takes 80% of global venture funding AI startups for the first time captured half of global venture funding in Q4 2024. Since then, that percentage has hovered around 50% — until Q1 of this year, when OpenAI’s record-setting round, along with a small handful of other enormous deals, pushed AI’s share to 80% of the quarterly funding total. Top 4 vs. everyone else It wasn’t just that AI as an industry captured the lion’s share of venture funding last quarter. Just four companies took nearly two-thirds of the entire pie, Crunchbase data shows. Four of the five largest venture rounds ever recorded were closed in Q1 2026, with frontier labs OpenAI ($122 billion), Anthropic ($30 billion), xAI ($20 billion) and self-driving company Waymo ($16 billion) collectively raising $188 billion, or nearly 65% of global venture investment in the quarter. Deal count falls even as dollars surge And while last quarter set an all-time record for venture dollars invested, more money went to fewer companies, continuing an overall downward trend for deal count we’ve seen since the beginning of 2021. That was the case not just in North America, where dollars invested surged 190% year over year, even as deal count dropped 26%. It also held true in Europe and Latin America. Only Asia saw a modest 5% bump in deal count along with more dollars invested. Related Crunchbase queries: - Global Venture Funding In 2026 - Global Venture Funding To AI Startups In 2026 - North America Venture Funding In 2026 Related reading: - Q1 2026 Shatters Venture Funding Records As AI Boom Pushes Startup Investment To $300B - North America Q1 Funding Surges Across Stages To Record Level - Global Investors Help Boost Latin America’s Late-Stage Funding Boom In Q1 - China Leads Asia’s Startup Funding To Its Highest Level In More Than 3 Years - AI Drives Europe’s Second Straight Quarter Of Funding Gain As Deal Volume Falls Sharply Illustration: Dom Guzman Stay up to date with recent funding rounds, acquisitions, and more with the Crunchbase Daily. ### Steve Bannon backs Anthropic over the Pentagon while calling the House for Republicans **Source:** Semafor **Link:** https://www.semafor.com/article/04/16/2026/maga-commentator-steve-bannon-predicts-republicans-will-hold-the-us-house The News Prominent MAGA commentator Steve Bannon on Thursday offered a rosy view of the populist right’s influence — and of Republicans’ fortunes, predicting they would keep the US House this fall. “I feel better than ever,” Bannon told Semafor World Economy in Washington, DC. He touted the success of conservative Senate candidate Ken Paxton in Texas and polls showing that less than a majority of South Carolina voters approve of GOP Sen. Lindsey Graham, whose hawkishness has made him a frequent political punching bag for Bannon. Bannon acknowledged that, while populist- and nationalist-leaning conservatives such as himself “were ascendant” in the early months of President Donald Trump’s second term, recent months have seen his wing’s interest “maybe not totally aligned” with the rest of the party. Among the issues on which his camp has split from the White House are the war with Iran — Bannon declared that “we’ve got to get out of the Middle East” — and mass deportations, which Trump’s aides have urged the party to downplay heading into the midterm elections after a remarkably aggressive start to immigration enforcement. But Bannon, a longtime Trump ally, vowed that “we will be ascendant again,” predicting that Republicans are “going to hold the House” majority. (The majority of polls this spring show that Democrats, not Republicans, are on track to claim control of the chamber next year.) One Republican Bannon was less bullish on: Vice President JD Vance. Asked if Vance was the “heir apparent” to Trump’s base for 2028, Bannon said that he continues to advocate for Trump to seek a third term in office despite the Constitution’s prohibition on such a move. Know More Bannon also underscored his sharp criticism of “out-of-control” artificial intelligence companies, saying that the AI startup Anthropic “had it right” in its feud with Trump’s Pentagon over the use of its technology in autonomous weapons systems and the surveillance of Americans. For populist conservatives, Bannon said, AI is “right next to immigration, our top issue. People are galvanized by this.” He called for “some modicum of regulatory control” over technology that he warned would hurt children and take away jobs. And he warned that congressional Republicans would keep trying to pursue legislation that would effectively nullify state-level attempts to regulate AI, despite having failed to do so in the past. ### STAT News warns voice-first chatbots will make AI's mental health problem worse, not better **Source:** STAT News (AI) **Link:** https://www.statnews.com/2026/04/16/voice-chatbots-ai-psychosis-mental-health/?utm_campaign=rss A Florida father recently sued Google after his son, Jonathan Gavalas, died by suicide following months of interaction with the company’s artificial intelligence chatbot Gemini. The case has rightly focused attention on how chatbots apparently reinforce delusions and foster emotional dependency. Yet, there is a critical detail easy to dismiss. Jonathan Gavalas was not just typing to Gemini. He was talking to it using Gemini Live, Google’s voice-based conversational mode. That distinction matters far more than the current debate acknowledges. Every week, around 800 million people interact with ChatGPT. According to OpenAI, roughly 0.07% of those weekly users show possible signs of psychosis or mania during their conversations, while 0.15% display indicators of suicidal planning or intent. Even if these figures are imprecise, they imply that hundreds of thousands of people worldwide who experience serious psychological distress interact with an AI chatbot. Most of those numbers come from the era of text. The shift to voice has just begun, and it will likely make things worse. Tech companies are racing to put AI chatbots in our ears. OpenAI is developing a dedicated voice-first device. Meta already offers smart glasses with built-in microphones and speakers that enable AI conversation. Apple supposedly plans to extend its AirPods for voice-based chatbot interaction. That makes the direction very clear: The primary way humans communicate with AI is moving from typing and reading to speaking and listening. For most users, this will feel like a convenience. For vulnerable people — those prone to psychosis, mania, depression, or loneliness — it may represent a serious and unexamined risk. In a recent Acta Neuropsychiatrica editorial, psychiatrist Søren Østergaard and I outlined why that is the case. Voice is how humans first learn language. Long before a child reads a single word in school, their brain is already wired to process speech. They naturally respond to tone, rhythm, emphasis, and emotional inflection. Text strips all of that away. When you read a chatbot’s response on a screen, there is an inherent distance because you are processing symbols, not hearing a humanlike voice. That distance creates natural cognitive barriers. You pause. You reread. You push back. Voice removes those barriers. Speech recognition is significantly faster, nearly three times as fast as typing. It is more seamless and far more emotionally engaging. When an AI speaks to you, it activates something deeper and older than literacy. This is not merely theoretical. A preprint of a randomized controlled study co-authored by OpenAI researchers found that people spent significantly more time interacting with voice-mode ChatGPT than with the text version, suggesting greater engagement. Voice initially appeared to boost certain positive outcomes, such as reduced loneliness. However, longer engagement with voice-based chatbots was linked to more negative psychosocial effects, inc ### Zvi Mowshowitz breaks down Claude Code's Auto Mode in Agentic Coding #7 **Source:** Zvi Mowshowitz **Link:** https://thezvi.substack.com/p/claude-code-codex-and-agentic-coding Claude Code, Codex and Agentic Coding #7: Auto Mode As we all try to figure out what Mythos means for us down the line, the world of practical agentic coding continues, with the latest array of upgrades. The biggest change, which I’m finally covering, is Auto Mode. Auto Mode is the famously requested kinda-dangerously-skip-some-permissions, where the system keeps an eye on all the commands to ensure human approval for anything too dangerous. It is not entirely safe, but it is a lot safer than —dangerously-skip-permissions, and previously a lot of people were just clicking yes to requests mostly without thinking, which isn’t safe either. Table of Contents Huh, Upgrades Claude Code Desktop gets a redesign for parallel agents, with a new sidebar for managing multiple sessions, a drag-and-drop layout for arranging your workspace, integrated terminal and file editor, and performance and quality-of-life improvements. There is now parity with CLI plugins. I can’t try it yet as I’m on Windows, aka a second class citizen, but better that then using a Mac. Daniel San is a fan and highlights some other features. Claude Cowork can connect to TurboTax or Aiwyn Tax and Claude can do your taxes for you, at least if they’re insufficiently complex. I’m filing for an extension, primarily because I’m missing some necessary documents from an investment, but also because think how much better Claude will be at filing your taxes six months from now. Claude Code now has full computer use for Pro and Max plans, for now macOS only. Claude Code auto-fix in the cloud allows Web or Mobile sessions to follow PRs, fixing CI failures and address comments to keep PRs green. Claude Code Dispatch can set its permission level. They recommend Auto, if available. Alex Kim goes over some features within the Claude Code revealed via the source leak. Anthropic offers the option to use Sonnet or Haiku as the end-to-end executor of your API agentic request, but to use Opus as an advisor model when there is a key decision. They suggest running it against your eval suite. An obvious follow-up is, are they going to bring this to Claude for Chrome or to Claude Code or Cowork? On Your Marks GPT-5.4-High reward hacks in the METR test and got caught. Accounting for this they get a disappointing time estimate of 5.7 hours to go with the misalignment issue. If you allow the hacks you get 13 hours, versus 12 hours for Claude Opus 4.6. Epoch, in cooperation with METR, proposes a new benchmark, MirrorCode, which checks the most complex software an AI can recreate on its own. Epoch AI: What are the largest software engineering tasks AI can perform? In our new benchmark, MirrorCode, Claude Opus 4.6 reimplemented a 16,000-line bioinformatics toolkit — a task we believe would take a human engineer weeks. This is a good illustration of ‘as AI improves it jumps rapidly from unable to do a given task to being able to consistently do a given task.’ What this cannot do for now is compare models from differen