Everything That Happened in AI Today Thursday, July 16

Kimi K3 arrived with 2.8 trillion parameters, one million tokens of memory, and the subtle energy of a new neighbor parking a battleship in the driveway.

Welcome to the Around the Horn Digest, where we track every AI story worth knowing so you can sound dangerously informed without opening 100 tabs. Moonshot's Kimi K3 dominated the model conversation, while Apple cleared a path into China, Google fought Europe over Android access, and DeepMind expanded its biosecurity work as a safety researcher reportedly resigned over a Pentagon deal. Nvidia and Japan built national physical-AI infrastructure, Fireworks became a $17.5B company, and power demand kept turning chatbot progress into a grid-planning problem. Meanwhile, agents learned to order DoorDash, use 1Password without seeing your password, build presentations, and review code at adjustable effort levels. At this rate, the only remaining human task is approving the calendar invite. Let's get into it.

Around the Horn - Thursday, July 16, 2026

The day's biggest release was Moonshot AI's Kimi K3, a 2.8T-parameter open model with native multimodality and a 1M-token context window. Moonshot says new attention techniques make decoding up to 6.3x faster and training roughly 25% more efficient, with open weights scheduled for July 27.

The launch was not impressive only because the number is large. Benchmark breakdowns put K3 near or above GPT-5.6 Sol Max and Fable 5 Max on several agentic tests, while Artificial Analysis said it leads multiple automation benchmarks and uses fewer output tokens than K2.6. Simon Willison found the model capable but expensive and token-hungry, a useful reminder that open weights do not automatically mean cheap operation.

The wider reaction was strategic. Thinking Machines' Tzachor Shaked said Chinese labs are no longer clearly behind public frontier systems, while Qu Xiaoyin argued enterprise customization around open models could become a larger business than selling one general-purpose chatbot. The open-model race is no longer the discount aisle. It is becoming the main floor.

🏆 TOP 5 NEWS

Google DeepMind, OpenAI, and Anthropic converged on pre-release testing, a U.S.-led regulator, and narrow rules for catastrophic-risk models. The safety push also favors labs that can afford expensive compliance, making regulation a market-structure fight as well as a risk-management one.
Commission provides guidance to Google for AI interoperability on... issued binding DMA measures requiring Google to open Android features to rival AI assistants and share anonymized search data with eligible competitors.
TSMC second-quarter profit spikes over 77%, topping estimates reported Q2 profit rose 77% on $40.2B revenue as TSMC added another $100B to its Arizona investment plan.
Japan Government, Industrial Leaders and NVIDIA Launch the World’s... announced a 140 MW national Vera Rubin AI factory with 13,750 Vera CPUs and 27,500 Rubin GPUs for physical AI, robotics, and open multimodal models.
Fireworks Secures $1.5 Billion in Series D Funding raised $1.505B at a $17.5B valuation after crossing $1B in annualized revenue and 40T daily tokens.

Honorable Mentions

Musk's xAI sues Grok user over sexualized 'deepfakes' sued a Grok user accused of using the model to create sexualized deepfakes and child sexual abuse material, seeking damages and a permanent ban.
Hack Reveals Suno AI Music Generator Scraped YouTube, Deezer, and... reported hacked Suno code showed scraping from YouTube Music, Deezer, Genius, podcasts, stock-music libraries, and other sources used for training.
SpaceXAI reported SpaceXAI is struggling with departures, operational disorder, and an inconsistent strategy as it tries to make Grok competitive with Claude.
The Oversight Board Says Leading AI Models Might Be Restricting... warned leading models may apply restrictive speech rules globally based on strict jurisdictions, limiting political criticism and protest-related expression.
Twenty-nine countries sign agreement to establish global AI... reported 29 countries signed an agreement to create the Shanghai-based World AI Cooperation Organization for international AI governance.

🍪 TOP TREATS TO TRY

1Password for Claude gives Claude credentialed access to websites and one-time passcodes while keeping secrets encrypted, scoped, biometrically approved, and invisible to the model.
NotebookLM is now Gemini Notebook renamed NotebookLM to Gemini Notebook and added deeper Gemini integration plus a secure cloud computer that can run code for grounded analysis.
You can now build AI agents that order DoorDash directly from the... covered DoorDash's dd-cli beta, which lets developers and agents search stores, build carts, and place orders from a macOS terminal.
Google Vids gets powerful upgrades with Gemini Omni added Gemini Omni video generation and editing plus personal avatars built from a selfie and voice recording.
Claude Code's /code-review effort levels rewrite the review strategy at every setting: Low runs a fast single-pass diff review, Medium adds context and multi-angle passes, High uses fresh-context subagents, and Ultrareview launches a cloud fleet that independently reproduces findings before reporting them. It supports branch diffs and GitHub PRs, delivers results as notifications, and includes three free Pro / Max runs before usage credits, typically about $5-$20 depending on change size.
GitHub open-sourced a text-space optimizer that trains reusable natural-language agent skills through trajectory edits, validation gates, and deployable best_skill.md files.
GitHub open-sourced a presentation system where any agent can build responsive React decks with charts, timelines, 3D scenes, animation, live data, and presenter tools.

🏢 Big Tech & Major Companies

Google's AI Mode now lets you link and interact with select apps expanded AI Mode from answering questions into taking actions across connected apps such as Instacart, Canva, and YouTube Music.
About Netflix reported 97B viewing hours in the first half of 2026, with non-English content accounting for more than a third of all viewing.
Netflix says around 300 titles used generative AI said roughly 300 Netflix titles used generative AI for post-production, crowds, battle scenes, and worldbuilding to cut time and cost.
Build Without Limits on Roblox introduced Build, a mobile-first creation system that turns prompts into playable Roblox games and 3D scenes before handing projects to Studio.
Alphabet shares fall on report its Gemini 3.5 Pro AI model is delayed reported Alphabet shares fell after Gemini 3.5 Pro's broader rollout slipped while Google worked to improve coding performance.
Alerting Parents if Teens Show Signs of Distress in Conversations... began alerting supervised parents when a teen discusses suicide or self-harm with Meta AI, after automated flags and human review.
Simon Willison reported xAI's Grok CLI uploaded full directories to Google Cloud, prompting a rollback, data deletion, default-retention changes, and an Apache 2.0 open-source release; The Decoder emphasized that the open-source release followed the data-exposure backlash.
NVIDIA and Japan Bring Full-Stack AI and Robotics to Every Industry showcased Japan partnerships spanning robotics, drug discovery, manufacturing, health care, automotive, and quantum computing.
Exclusive reported violent threats and real-world security incidents against AI executives are rising alongside public backlash.
@AndrewCurran_ Andrew Curran noted Alphabet closed down 4.4% and SpaceXAI down 3.1% as the market reacted to Gemini 3.5 Pro delays and broader AI-stock pressure.
@sama Sam Altman said OpenAI did not have its best previous 12 months, took much of the blame, and promised a stronger year centered on user freedom, agency, and wealth creation.

Apple Intelligence received Chinese approval through Alibaba's Qwen and Baidu integrations, giving Apple a path to launch its delayed AI features in China; The Information reported the service was among seven newly approved systems, mostly from smartphone makers.
The EU ordered Google to give rival assistants comparable Android access, with implementation extending into July 2027 while Apple's upgraded Siri still lacks a European timeline.
Google DeepMind launched a bioresilience program with Isomorphic Labs for pathogen surveillance, vaccine design, and outbreak response.
Google DeepMind safety researcher Alex Turner reportedly resigned over the company's classified Pentagon deal after roughly 600 employees signed an internal protest.
AI data-center demand pushed up prices across PJM, the largest U.S. grid, intensifying backlash against new projects.
Whatnot acquired Shaped to improve real-time search and recommendations for live shopping.
Truth Social plans to license real-time platform data to Wall Street and is discussing AI-training deals, following X and Reddit into the data-licensing business.
AWS is ending sales of two cloud-hardware products for private data centers as providers rethink where AI and edge workloads should run.

💼 AI Productivity, Labor & Economics

CVS is building a Google-powered assistant for scheduling checkups, refilling prescriptions, and checking insurance coverage.
AMI Labs founder Alexandre LeBrun rejected AGI and superintelligence labels, framing the company's work around practical agent behavior instead of frontier-lab hype.
The Verge argued many AI-generated film projects function more like direct-to-video marketing for production workflows than competition for human-made movies.
Meta employees sued over claims that AI-generated layoff lists disproportionately targeted workers with disabilities or protected leave, which Meta denies.

🤖 AI Agents, Workflows & Developer Tools

Mem0 became an official Anthropic connector, so Claude users can add cross-session memory from the Connectors directory without a CLI, config file, or manual MCP installation. Once connected, chats on claude.ai, Desktop, and Cowork can carry context across weeks instead of resetting every session.
Delba Oliveira, via Claude Developers, explains how to choose among turn-based agent prompts, goal loops with verifiable success criteria, scheduled loops, and proactive routines. She recommends explicit stop conditions, self-verifying skills, fresh-context review agents, right-sized models, and small pilots before letting long-running workflows loose.
@thsottiaux OpenAI engineer Thibault Sottiaux said GPT-5.6 Codex occasionally deleted files in full-access mode, often after changing the HOME environment variable or making an honest mistake; OpenAI is adding safer defaults, harness protections, and a detailed postmortem.
Andreessen Horowitz backed an agent-safety startup reported Andreessen Horowitz backed a startup that 'parents' agents by limiting file access, actions, and spending authority.
@KimptonAI Kimpton AI introduced a firm-wide research brain that reads proprietary memos, models, notes, and archives to power signals, backtests, coverage, and analysis that improve with every file added.
Expanded SkillOpt Ablations, Skill-Aware Reflection, and... reported skill-aware reflection, memory consolidation, and validation-gated updates improved SkillOpt's stability, while SkillOpt-Sleep optimized skills offline with bounded downside risk.
@btaylor Bret Taylor highlighted Inkling as Thinking Machines' first open model and a bet that companies will want customizable multimodal intelligence rather than one-size-fits-all APIs.
@GergelyOrosz Gergely Orosz examined the growing gap between spectacular AI demos and the harder engineering work required to make agents dependable inside real software teams.
@anton_lozhkov Anton Lozhkov shared implementation details and developer reactions around open models and the infrastructure needed to run them efficiently.
@emilkowalski Emil Kowalski showed how careful interaction design can make AI interfaces feel faster and more legible, emphasizing product polish over raw model capability.
@cerebras Cerebras highlighted high-speed inference as a practical advantage for agent loops, where every extra second compounds across many model calls.
@Suhail Suhail Doshi argued the current wave of model releases is shifting product advantage toward teams that can turn frontier capability into simple workflows quickly.
@Steve_Yegge Steve Yegge argued coding agents are changing software work from writing every line to supervising parallel agents, raising the premium on specification, review, and taste.

Sable raised $45M for Aidan, a computer-using sales engineer that gives live demos, answers questions, and handles onboarding for customers such as Notion and Decagon.
OpenAI's Codex Micro is a $230 Work Louder keyboard with agent-status keys, command shortcuts, a workflow joystick, and a reasoning dial; The Decoder said the hardware targets the window-hopping involved in supervising agents.
Vint Cerf joined Innovation Labs as it develops DNSid, a domain-linked identity standard meant to make autonomous agents auditable on the open internet.
Oak emerged from stealth with $60M for an identity control plane that cleans up permissions across employees, software, and agents.
OpenAI's Codex now encrypts agent-to-agent instructions for larger GPT-5.6 models, making internal delegation harder for developers to inspect.
Google's Gemma 4 received faster Hopper GPU performance, tool-calling fixes, and fewer truncated responses without changing its model name.

🔬 Models, Research & Benchmarks

@zxytim Zxy Tim said his team is prioritizing the intelligence frontier over video generation because video models are useful products but do less to advance core reasoning capability.
Latent Thought Flows with Text Compression presented research on compressing latent thought flows so models can reason more efficiently without emitting every intermediate step as text.
@twid Tim Dettmers highlighted latent-thought compression as a route to more efficient reasoning by allowing models to move and compress internal thought representations instead of spelling every step out in text.
@asterailabs Astera Labs highlighted infrastructure work aimed at feeding faster accelerators with enough memory and connectivity, a reminder that model progress increasingly depends on the systems around the chips.
@googlegemma Google Gemma announced Gemma 4 31B on LiveKit Inference as a voice-agent default with GPT-4.1-level answer quality, 192 ms time to first token, and about 6x lower cost.
Blind Video A/B runs blind A/B tests asking which generated video looks more physically plausible, helping researchers measure failures on chained physical events.
[2607.13031] The Seriality Gap in Video Diffusion Models found bidirectional video diffusion models break down as tasks require longer chains of dependent events because extra denoising steps do not create scalable serial computation.
The Seriality Gap in Video Diffusion Models showed autoregressive generation and deeper backbones close the 'seriality gap' by adding explicit step-by-step computation.
GitHub released the official PyTorch code for training and evaluating seriality-gap experiments.
[2511.20820] SAGE introduced SAGE, an agentic framework that interprets sparse-autoencoder features through hypotheses, experiments, and validation rather than one-shot labels.
[2605.03808] Agentic-imodels showed agentic interpretability tools can evolve themselves through automated research and improve downstream explanation performance.
[2606.24026] Can Language Model Agents be Helpful Circuit... tested whether language-model agents can explain neural circuits through observation, hypothesis generation, and validation loops.
@fnruji316625 a research thread framed SAGE, Agentic-iModels, and HyVE as an emerging agentic-interpretability stack where models generate hypotheses, run experiments, and validate explanations of their own circuits.
@JingyuanLiu123 jingyuan Liu highlighted Inkling's use of the Muon optimizer and modular-manifold constraints to keep weight norms controlled during training.
Modular Manifolds introduced a geometric framework for co-designing optimizers with manifold constraints, including manifold-Muon updates with better conditioning and predictable weight norms.
[2506.02285v1] Why Gradients Rapidly Increase Near the End of Training explained late-training gradient spikes as an interaction among weight decay, normalization, and the learning-rate schedule, then proposed a correction that also lowers loss.
@aaron_defazio Aaron Defazio's profile links to work explaining why gradient norms spike near the end of training and how correcting the interaction among weight decay, normalization, and learning-rate schedules lowers loss.
@tszzl Tzachor Shaked predicted open-weight models that run themselves, replicate workflows, and train improved behavioral versions are not far away.
@frank_ Frank Slootman discussed the growing premium on models and tools that deliver usable work rather than benchmark theater, reinforcing the shift toward production outcomes.
Inkling hosts Inkling, a 975B-parameter multimodal mixture-of-experts model with 41B active parameters that reasons over text, images, and audio.
Inkling by Thinking Machines now available on Modal made Inkling available with a custom DFlash speculative-decoding system for faster inference.

Thinking Machines Lab released Inkling, an open 975B-parameter mixture-of-experts model for enterprise customization; its launch post details 41B active parameters, a 1M-token context window, multimodal training, and fine-tuning through Tinker, while The Information noted architectural cues from DeepSeek V3.
OpenAI introduced GPT-Red, an automated red-teaming model that attacks other models to expose prompt-injection failures and helped harden GPT-5.6 Sol.
Bonsai 27B is a freely modifiable reasoning model compressed to run on iPhone-class hardware, with Apple reportedly testing the underlying method.
Sakana AI added Nvidia's Nemotron models to Fugu so specialized open models can work as coordinated agents behind one API; The Decoder framed it as a test of collective open-model intelligence against single frontier systems.
A University of Pennsylvania professor reportedly used GPT-5.6 Sol Pro to disprove a 30-year-old statistics conjecture in roughly 90 minutes.
Google Research introduced SensorFM, trained on more than one trillion minutes of wearable data from five million people and transferable to 35 health-prediction tasks.
Anthropic research found Claude's value patterns shifted across languages, including warmer Hindi responses and more rigorous Russian ones.

🤖 Robotics & Physical AI

New York school district is testing lifelike robot teachers reported a New York district is piloting Realbotix humanoid teachers for STEM support and 24/7 homework help while keeping humans in charge.
Tony Zhao on X Tony Zhao introduced ACT-2 Preview, a home-robotics model that learns a new behavior from one fine-tuning example and reached 99% success across 785 trials in unseen homes.
@mimicrobotics Mimic Robotics unveiled the tendon-driven M1 hand and U1 wearable exoskeleton, pairing human-range dexterity with direct demonstration capture for general-purpose manipulation.
Mimic Robotics explained Mimic's full-stack dexterity strategy: human-like hardware, wearable demonstration capture, a data pyramid, and custom middleware to avoid cross-robot learning gaps.
Sunday on X Sunday Robotics introduced ACT-2 Preview and its Memo home robot, claiming 99% autonomous success in unseen homes without collecting household-specific training data first.

🏛️ Policy, Safety & Society

Someone Used A.I. to Write an Unauthorized Biography of Me. I... documented an unauthorized AI-written biography filled with fabricated insights, part of a flood of low-quality automated books on Amazon.
Can AI Fix Bad Rules? showed an ambiguity rubric plus gradient-boosted trees predicted disputed contracts with roughly 75% accuracy, suggesting AI could grade and rewrite rules before conflicts occur.
@tobi Shopify CEO Tobi Lutke argued anti-AI criticism often overestimates human code quality, noting that frontier models can already clean up plenty of 'human slop.'

Anthropic is hiring for 32 roles focused on preventing catastrophic misuse across explosives, nuclear weapons, financial fraud, cybercrime, and other high-risk domains.
The White House left room for future action on open models as officials balance China-related risks against support for the U.S. open-source ecosystem; Semafor argued the problem extends beyond access because U.S. labs suspect distillation while businesses adopt cheaper Chinese systems.
Common Sense Media gave Google's AI Overviews and AI Mode its lowest child-safety rating after tests found homework completion, deepfake guidance, uneven crisis handling, and unreliable sourcing.
Australia announced a national Office of AI with planned standards, creator protections, and obligations for large data centers.

🛠️ AI Tools & Products

Reelful turns camera-roll clips and photos into scripted, voiced, captioned short videos for TikTok, Instagram, and YouTube Shorts; no pricing details.
A former Ultrahuman hardware VP raised $5.5M for devices designed to control agents instead of merely recording personal context; no pricing details.
Claude for Teachers gives verified U.S. K-12 educators free premium access, teaching skills, and standards-mapped curriculum support.
Apple's Siri AI public beta gives iOS 27 testers broad access to Apple's screen-aware, device-connected assistant before the fall launch.
Superhuman upgraded auto-drafts to write reply options from a user's prior tone and conversation context; available on paid plans.
Suno said a breach involved outdated code and no sensitive personal information after a hacker claimed its source detailed scraping from songs, lyrics, stock music, and podcast feeds.
Roblox Build turns mobile text prompts into playable game drafts, adding a creator-marketplace angle to Roblox's no-code push.
Overtone raised $18M for a voice-first dating service from Hinge founder Justin McLeod.
Waze added Gemini-powered destination search, conversational road reports, motorcycle routing, and more personalized navigation.

📊 Fundraising & Deals

Fora Raises $60M at $1 Billion Valuation, Defining a New Category... raised $60M at a $1B valuation for its platform supporting independent travel advisors with technology, training, and community.
Sheryl Sandberg leads $10 million investment in AI-powered vehicle... raised $10M led by Sheryl Sandberg for smartphone-based vehicle damage scanning used by fleets, auctions, lenders, and manufacturers.
Chinese filing implies DeepSeek valuation of around $52 billion reported a Chinese filing implied DeepSeek was worth about $51.8B.
Applied Computing lands $20M to expand foundation AI for energy raised $20M to expand physics-informed foundation models for energy plants, reliability, efficiency, and emissions work.
Bunkerhill Health raises $55 million to put AI agents to work... raised $55M to deploy hospital agents for scheduling, follow-ups, wait times, and paperwork through its Carebricks platform.
@Nim_Ravid1 Nim Ravid introduced Sable's Aidan as a computer-using AI employee for live demos and onboarding and announced a $45M raise from Sequoia and 8VC.

TSMC plans another $100B in U.S. investment as AI demand lifts its spending plan and revenue outlook.
Japan plans to develop homegrown models on Nvidia Vera Rubin chips, extending the sovereign-compute race.
microagi raised a $55M seed round to train factory robots on footage and sensor-glove data from more than 20,000 workers across 15 countries.
Elorian reportedly raised a $55M seed round at a $300M pre-seed valuation before launching a product.
Applied Computing separately detailed its $20M Series A and plant-wide model for oil, gas, and petrochemical operations.
Rime raised a $24M Series A for enterprise voice models trained on its own conversational data.
DeepSeek could file for an IPO as soon as this year while Zhipu's frontier push gives China's market a new post-DeepSeek storyline.

🎥 Video Interviews, Tutorials & Talks

Daniel Kokotajlo argues that the public AI-risk debate should focus less on “doomerism” and more on two concrete failure modes: companies losing control of superintelligent systems and a tiny group of executives or governments concentrating unprecedented economic, political, and military power.
This agent-workflow tutorial argues that companies will get real leverage from agents by moving them out of chat boxes and into shared task systems, where humans scope work, agents execute in parallel, and eval checklists decide whether outputs are good enough.
YC’s Eve Bouffard argues that as models get better at one-shotting software, the scarce skill shifts from execution to imagination: capturing your raw thinking in public, then using agents to turn that stream into tools, dashboards, websites, and research artifacts.
This Kimi K3 walkthrough argues that Kimi K3 matters less because it wins isolated benchmarks and more because frontier-ish intelligence is becoming distributable, cheap enough to reshape model economics, and no longer concentrated inside a few U.S. closed-model labs.
Weco’s Zhengyao Jiang argues that autoresearch agents will not replace human researchers so much as move them up the stack: humans design the evals, abstractions, and research environments, while agents execute, combine ideas, and search huge experimental spaces.
Demis Hassabis and Sergey Brin said the web is likely to become agent-first within a few years, with machines using the internet differently than humans do, while Brin guessed AGI arrives before 2030 and Hassabis placed it just after.
3Blue1Brown explains that cross-entropy, the loss function behind language-model training, is best understood as a compression problem: a model improves when its predictions make text less surprising and therefore easier to encode efficiently.
Matt Wolfe shows ChatGPT 5.6 as less of a chatbot upgrade and more of an operating layer that combines assistant work, browser control, coding, file actions, scheduled tasks, app connectors, and hosted sites into one agent workspace.
Jürgen Schmidhuber argues that today’s screen-bound language models are only one slice of AI progress, and the real inflection point comes when world models, physical robots, and self-replicating machinery bring AI out of text and into the physical economy.

💡 Industry Commentary & Analysis

Lydia Hallie, via Claude Developers, separates model choice from effort in Claude Code: the model determines what the system knows and can reason through, while effort controls how much context it reads, how deeply it plans, how often it verifies, and how long it persists. Her rule of thumb is to fix the prompt and context first, upgrade the model when it lacks capability, and raise effort when it simply did not work hard enough.
Addy Osmani argues that when agents automate the repetitive reps that once built engineering judgment, juniors must earn taste deliberately: read more code than they generate, keep a wrong log, build evals from real AI-generated pull requests, take one system end to end until it breaks, calibrate autonomy by task, and own the ungradeable final mile of choosing and verifying worthwhile work.
VentureBeat framed Kimi K3 as the largest open-source model yet and a direct challenger to leading U.S. systems; TechCrunch's pre-launch report captured expectations that it would close the gap with Anthropic's Opus line.
Google argued the EU's Android-interoperability and Search-data measures could weaken privacy and security guardrails for European users.
TechCrunch described DoorDash's dd-cli as an early move toward software designed for agents as well as humans.
TechCrunch emphasized that Vids can create a digital version of you from a selfie and voice sample, then generate and edit scenes through Gemini Omni.
Data Center Dynamics reported TSMC expects roughly four more Arizona fabs as it expands leading-edge logic and advanced packaging in the U.S.
CNBC connected Nvidia's Cosmos 3 Edge launch with a broader Japanese physical-AI push spanning factories, robotics, and industrial partners.
CNBC reported Fireworks diversified beyond its early reliance on Cursor as more companies sought cheaper specialized and open-model inference.
Politico placed Dario Amodei's $1M donation inside a widening political-spending fight between pro-regulation and anti-regulation AI groups.
TestingCatalog captured early Kimi K3 Arena outputs before launch, showing elaborate generations from the rumored multi-trillion-parameter model.
Ahmad Osman argued Kimi K3 felt unusually dense for a sparse model and highlighted the significance of free access plus forthcoming open weights.
Gavin Purcell speculated Kimi K3 could pressure public-market narratives once investors absorbed how competitive the model had become.
Moonshot AI announced Kimi K3's efficient attention systems, long-horizon coding focus, live web access, and July 27 open-weight release.
Ethan Mollick argued polished frontend demos are becoming the new model sycophancy because beautiful websites and 3D scenes win attention faster than deeper capability evidence.
LiveKit positioned Gemma 4 31B as a faster, cheaper voice-agent default with strong tool use, instruction following, and multi-turn coherence.
Bolt showed agents such as Claude Code and Codex generating interactive decks with bespoke layouts, animation, typography, live data, and working prototypes.
Sable described Aidan as a real-time conversational employee that watches the screen, navigates interfaces, and adapts live demos and onboarding to the person in front of it.

Axios argued electricity demand has become the clearest stress test for the AI boom as scaling runs into power, environmental, and financial constraints.
SoftBank founder Masayoshi Son predicted AI could generate 20% of global GDP by 2040 while requiring roughly three terawatts of data-center power.
Hugging Face CEO Clem Delangue said enterprises are moving more production workloads toward open and private models as cost, control, and ownership concerns squeeze closed APIs.
Meta's Adam Mosseri said companies may soon cap token budgets per engineer, turning internal AI use into a cost-management problem.
Anthropic's latest ad drew backlash for doomer imagery, surveillance scenes, and graveyard visuals while presenting Anthropic as the responsible lab.

Previous Around the Horn Digests

Catch up on recent digests:

That's a Wrap

That is 174 unique source links from one day. If you made it to the bottom, your context window is now legally considered frontier-scale.

For the daily version, make sure you are subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you do not have to.

See you tomorrow.

P.S. Know someone who would find this useful? Forward this to them and tell them to subscribe.

Everything That Happened in AI Today (Thursday, July 16, 2026)