The Neuron AI News Digest - Wednesday, October 16, 2025

‍Welcome to today's AI news digest.

Today's AI landscape in October 2025 showcases... 

• Apple dropped a new M-series computer to compete with NVIDIa's Spark.
• Google investing $15B in India's largest AI hub.
• Microsoft secured 200,000 NVIDIA GPUs for infrastructure.
• Apple launching the M5 chip with 3.5x faster AI processing.
• Meta partnering with Arm for energy-efficient chips.
• OpenAI forming safety councils amid regulatory scrutiny.
• Firefox adding Perplexity search.

...Plus dozens of AI startups raising millions for tools ranging from autonomous website builders and insurance automation to humanoid robots and edge computing, while research reveals global AI concerns, workplace productivity gains of 5.4%, and the transformation of industries from call centers to software development through AI-powered solutions.

The Big 🍎: Apple Unleashes M5: 4x Faster AI Performance Hardware

The Big Deal: Apple just announced the M5 chip, and it's a beast. We're talking 4x faster AI performance compared to the M4 chip that came out earlier this year.

What's New: The M5 packs a souped-up GPU with something called a "Neural Accelerator" in each core (fancy speak for: it runs AI tasks way faster), plus a beefier CPU, faster Neural Engine, and more unified memory bandwidth. Translation? Your MacBook, iPad, or Vision Pro is about to handle AI tasks like butter.

Why This Matters: As AI features become standard in everything we use, having hardware that can actually run these models locally (on your device, not in the cloud) is huge. Faster processing means smoother AI experiences without waiting for the internet.

Where You'll See It: Coming to the 14-inch MacBook Pro, iPad Pro, and Apple Vision Pro.

Around the Horn (Company News)

• Product launches
  
  
  • Google launched Coral NPU, an open-source AI platform for edge devices delivering 512 GOPS at milliwatt power levels.
  • ‍DirecTV and Glance announced AI-powered interactive screensavers launching in 2026.
  • Google added AI scheduling capabilities to Gemini that integrate with Gmail and Calendar.
  • Tech and crypto-focused Erebor Bank received conditional federal charter approval from the OCC.
  • Google added collapsible ads to search results and implemented AI Overviews for search result summaries.
  • Honor launched Magic 8 phones with huge batteries and extensive AI features.
  • Google Meet launched virtual makeup that applies realistic, motion-tracking makeup in 12 styles to instantly enhance your appearance during video calls.
  • Adobe had a suite of new announcements recently:
    
    • Adobe LLM Optimizer measures where your brand appears in ChatGPT and other AI chat results, then automatically fixes gaps so customers can find you when they search.
    • LLM Visibility Checker (Chrome extension) shows you what ChatGPT and other AI tools can actually see on any webpage you visit—free to try.
    • Audience Agent analyzes your CRM and website data to automatically build lists of decision-makers at companies most likely to buy your product.
    • Journey Agent creates multi-channel marketing campaigns across email, web, and mobile, then optimizes them by finding where customers drop off.
    • Data Insights Agent lets your sales and marketing teams ask questions about customer data in plain English and get instant charts and forecasts.

• Fundraising
  
  • Flint secured $5M seed funding for its AI website automation platform, with early adopters reporting 50% higher Google Ads conversions.
  • MIT spinout Vertical Semiconductor raised $11M to develop vertical GaN chips for AI data centers.
  • Eightfold AI co-founders raised $35M to launch Viven, creating AI digital twins that retain employee knowledge when colleagues are unavailable.
  • Rhoda AI raised $162.6M Series A for humanoid robots while Genesis AI secured $105M seed funding for similar technology.
• Partnerships
  
  • Meta partnered with Arm to use energy-efficient chips for AI systems across platforms starting 2025.
  • Microsoft and Nscale signed a deal for 200,000 NVIDIA GB300 GPUs across US and Europe.
  • Firefox added Perplexity AI as a global search engine option, providing AI-generated answers with citations.
  • Salesforce expanded partnerships with OpenAI and Anthropic, integrating their AI models into Agentforce 360 platform.
  • Insurance AI startup Liberate raised $50M Series B at a $300M valuation, with clients reporting 15% sales increases and 23% cost reductions.
• Updates
  
  • Coco Robotics established a physical AI lab led by UCLA Professor Bolei Zhou as Chief AI Scientist.
  • Google invested $15 billion to build India's largest AI hub in Visakhapatnam.
  • Salesforce says it saved $100 million annually using its Agentforce AI in customer service, with over 12,000 customers now using the platform.‍
  • Oracle Cloud will deploy 50,000 AMD AI chips as an alternative to Nvidia.
  • South Korea decided to “roll back” their AI textbook program, stripping the books of their official textbook status, after only four months due to inaccuracies and data risk.

Treats to Try (Companies/Tools)

• Docker Model Runner lets you run AI models locally on your computer using Docker, with no complex setup needed and full GPU support for faster performance—free to try.
• Wispbit enforces your team's coding standards through automated code reviews that flag specific violations in any environment—free to try.
• Traycer transforms your coding workflow by breaking complex tasks into manageable phases, then generates code for each phase while keeping you in control—free plan available, then $10/month for Lite or $25/month for Pro.
• Waydev AI shows you exactly how your engineering team is performing with real-time metrics on code quality, project progress, and developer productivity—paid only ($649/year).
• Emergent.sh turns your text descriptions into fully functional web apps in minutes—just type what you want to build and it creates everything from customer dashboards to e-commerce sites with no coding required—free trial, then paid only.
• Strawberry Browser lets you build AI companions that autonomously find leads, research companies, and draft emails for you without manual intervention—free to use while in closed beta.
• CometChat speeds up your AI agent deployment from months to days by providing ready-made chat infrastructure, freeing you to focus on what makes your agent unique—free for first 500 customers until year-end.
• Metorial lets you connect your AI models to thousands of APIs with one line of code, so you can build agents that book flights, search databases, or call external tools without complex integration work (YC F25 batch).
• KaneAI lets you test software with simple English commands across thousands of real devices, eliminating coding requirements.
• Supercut.ai automatically edits your long videos into shareable clips and suggests social media posts for each one—no editing skills required.
• Riverside's Co-Creator turns your recordings into ready-to-publish clips, thumbnails, and social posts with a single prompt, all from within your project dashboard.
• Scriber Pro transcribes your 4.5-hour videos in just minutes on your Mac, completely offline with no file size limits.
• 10Web turns your business description into a complete WordPress website in minutes, with built-in hosting, 90+ PageSpeed scores, and no coding needed—$10/month.
• Xona.ai speeds up your interior design workflow by turning reference photos into custom designs with instant restyling, decluttering, and team collaboration features—paid only ($9-$99/month).
• Osaurus lets you run AI models locally on your M-series Mac with built-in model management and OpenAI compatibility.
• Meilisearch lets you add ChatGPT-like search to your app with features like "search by image" and location-based filtering while maintaining lightning-fast response times.
• Flask provides video collaboration tools for creative teams.
• Trott organizes your saved reels and shorts with AI.‍
• Campfire automates your month-end accounting close and multi-entity consolidation, helping finance teams close their books in hours instead of days, w/ customers reporting 70-80% faster closes. (raised $100M).‍
• Finch automates paralegal work for your personal injury cases, answering intake calls, retrieving medical records, filing claims, and writing demand letters so you can double your caseload without hiring more staff (raised $20M).
• The E-MM1 claims to be the world's largest multimodal dataset for advancing AI research.

Intelligent Insights (Technical/Deep Dives/Perspectives)

• LMSYS published their review of NVIDIA's new DGX Spark, a compact AI supercomputer with 128GB unified memory and Blackwell GPU that runs large language models locally.
• Alex Kantrowitz writes that AI has a "sameness" problem, producing the "average of averages" and that this issue needs to be resolved before AI has a chance at sticking around past its novelty stage.
• The disconnect between tech and traditional business becomes clear in this eye-opening essay from Boyd Kane on why your boss isn't panicking about AI, revealing how business leaders evaluate AI through practical lenses of cost-benefit rather than theoretical capabilities.
• BoldVoice created a fascinating AI accent analysis that maps the world's English accents in 3D space using 30 million speech recordings, revealing how accents cluster through geographic proximity and historical migration patterns.
• UK unemployment hit a four-year high according to this sobering labour market report showing 4.8% unemployment with 2.4 unemployed individuals per vacancy.
• Tracking real-time install trends is becoming a competitive advantage for software vendors according to this data-driven analysis of AI coding tool adoption patterns.
• Modern LLMs have quietly mastered character-level operations like string reversal, according to this deep dive from Tom Burkert, effectively giving non-programmers powerful automation capabilities.
• Unsupervised AI agents burned $200 in just two hours reveals this sobering experiment showing the need for strict spending limits and human oversight.
• Gregory Barber writes for Wired how AI is rewiring the very nature of programming, and how writing assembly code might be a truer path to better AI outputs.
• Jampa Uchoa writes that large language models remain far from 99% accurate, with users' tendency to blindly trust AI outputs emerging as a critical risk factor, but using AI to trivialize hard problems is the winning product strategy that not enough companies are implementing.
• Not a single nation had more people excited than concerned about AI according to this global AI attitude survey across 25 countries by Pew Research Center.
• OpenAI subpoenaed seven nonprofit organizations critical of its transition raises serious questions in this troubling development about how tech companies respond to criticism.
• AI is reshaping India's massive call center industry where companies can now automate the work of 15 agents for just 100,000 rupees monthly.
• Hardware hackers jailbroke the "bricked" Humane AI Pin, showing that physical access to even well-secured IoT products remains the ultimate vulnerability.
• Technological optimism is warranted, argues this analysis from Import AI, estimating less than a 1% chance AI won't deeply transform our economy.
• Karan Sharma says AI is enabling "home-cooked software", where non-coders create custom tools in hours instead of weeks, and that AI will turn everyone into a software engineer long before it replaces them.
• Giles Thoma's journey building an LLM from scratch over 22 detailed posts culminates in this payoff moment when loading pre-trained GPT-2 weights transformed rough output into coherent responses.

Here's the X posts that caught our eye:

• TestingCatalog spotted an in‑app banner indicating an upgrade to Gemini 3.0 Pro.
• Omar shared a demo where Claude Code subagents coordinate multi‑step coding work.
• VraserX posted hands‑on clips and panned Veo 3.1 for artifacts and inconsistency.
• neural_avb demoed an agentic memory prototype that persists context with vector search and tools.
• A thread introduced the Encode‑Think‑Decode method, a recursive latent‑thoughts approach that  scales test‑time reasoning (paper)..
• Akhaliq highlighted the DeepMMSearch‑R1 method, an Apple take on multimodal retrieval‑augmented search (paper).
• A summary claimed a small domain‑trained LLM found a novel way to make tumors more responsive to immunotherapy.
• TestingCatalog said Google is preparing a Gemini Agent prototype for multi‑step web tasks.
• Sayash Kapoor launched HAL agent‑eval stack that spins up many sandboxed VMs to test reliability (paper).
• Assaf Shocher posted a short explainer on linearity in neural nets and why it matters.
• Omar Khattab dissected how a recursive LM decides steps using peeking, map‑reduce, and summarization.
• VraserX highlighted AI Scientist‑v2 agent that proposes experiments, analyzes data, and drafts papers (paper).
• Virat shared an AI hedge‑fund stack combining specialized agent workflows with open repos to reproduce.
• AISafetyMemes share how the majority of articles are now written by AI.
• Evil Martians showed how agentic traces capture plans, tool calls, and retries—replayable via OpenTelemetry.
• Alex L. Zhang teased that context window scaling could be simpler than expected, with details forthcoming.
• Guillermo Rauch praised Google’s Nano Banana model as fast and high‑quality.
• Millie Marconi summarized Meta’s Early Experience training—no rewards, demos, or supervision—sometimes outperforming both (paper).
• Dimitris Papailiopoulos shared 1,700 memory experiments on allocating weights, KV cache, and test‑time compute (paper).
• Rohan Paul spotlighted the Trading‑R1 model, a 4B model that writes  theses and turns them into trades (paper).
• Rauch showed ChatGPT Apps can run Doom in‑chat by embedding a Next.js app via an MCP tool.
• Khattab noted a PhD student built the recursion demo in four weeks—signaling the research clip.
• Lenny Rachitsky argued Claude Code’s best use is a thinking partner for PMs and operators.
• Junde Morsen Wu summarized a General Adaptive Attack framework that iteratively bypasses guardrails (paper).
• Sundar Pichai said the C2S‑Scale 27B model generated a cancer hypothesis later validated in living cells.
• Rohan Paul highlighted agents that learn “experience rules” in‑prompt to boost performance without retraining.
• CDT reported school AI adoption is linked to increased risks like breaches, bias, and harmful chatbot interactions.
• Rauch also pointed to a v0‑made playground used for Nano Banana.
• Agent Learning via Early Experience (paper).

What videos we’re watching atm:

• The origin story of Nano Banana (and how its now a marketing case study lol).
• Software is eating labor, and this video explains how.
• Anthropic researcher Sholto Douglas on Claude Sonnet 4.5 and the “myth” of the AI scaling plateau.
• Peter Yang interviewed a developer named Ryan Carson who shares his 3-step system to build production apps w/ AI live in action (its this one: AI Dev Tasks repo, which he uses w/ Ghostty and Temux).
• Muhammad Zain breaks down how to use the new Claude Code with Cursor or VS Code to build a Snapchat-style face filter app live.
• If you’re a coder using Claude Code, and you don’t know about Claude’s “thinking” levels, this short video explains it well.
• Here's a video we covered on the website! AI Isn't Slowing Down: Nathan Labenz of The Cognitive Revolution on Real Progress, Agents, and What's Next.

One to Watch 👀 

Poolside Is Building a 2-Gigawatt "AI Superbrain" in West Texas

THE NEWS: AI coding startup Poolside just announced Project Horizon—a massive self-powered data center in West Texas that'll eventually generate 2 gigawatts of electricity. That's roughly equal to the Hoover Dam's output, all dedicated to training AI.

THE SETUP: The facility will sit on 500+ acres of the Mitchell family's Longfellow Ranch in the heart of the Permian Basin—America's busiest oil field. Instead of relying on Texas's increasingly strained power grid, Poolside will tap natural gas on-site to generate its own electricity.

CoreWeave (Nvidia's favorite cloud partner) is providing 40,000+ NVIDIA GB300 GPUs starting December 2025 and will anchor the first 250MW phase coming online late 2026, with an option for 500MW more.

THE TENSION: Why does a two-year-old AI startup need Hoover Dam-level power?

Because Poolside co-CEO Eiso Kant believes "to compete at the frontier you need to be vertically integrated from dirt to intelligence."

HERE'S THE BIGGER PICTURE:

When we interviewed Poolside CEO Jason Warner earlier this year, he laid out his thesis: "Intelligence on compute" is just 36 months away. His bet is simple—whoever scales compute fastest wins AGI.

Poolside's secret weapon is what Warner calls "reinforcement learning via code execution feedback." Their AI doesn't just predict code—it writes it, runs it in real containerized environments, watches what breaks, and learns from failures. It's trial-and-error learning at massive scale.

Warner told us: "The holy grail is to get a thought process dataset"—capturing not just inputs and outputs, but the reasoning in between. Programming provides that missing link because every line of code is testable.

Key to this initiative: they're betting that reinforcement learning—not next-token prediction—will soon become the largest line item on their compute budget.

HOW THEY'RE DOING IT:

When we chatted with Eiso earlier this year, he explained Poolside's built an environment with 800,000 GitHub containers—two orders of magnitude larger than anyone else—where AI agents write code, run it, fail, learn, and try again.

Kant told us they're "the largest Kubernetes users in the world" with 30 million container images. Their agents operate in three modes: real-time pair programming, async task execution, and fully autonomous operation (monitoring logs, watching CI pipelines, triaging bug reports).

Jason Warner calls it "reinforcement learning via code execution feedback"—and it's patterned after AlphaGo. Start with simple tasks. Gradually increase complexity. Always keep the AI at the edge of what it can handle so it learns through experiential trial and error.

THE KEY INSIGHT: Most AI companies train on static datasets. Poolside generates synthetic data by having AI solve millions of real coding tasks—and 70% of their budget goes to this exploration phase.

Warner explained: "The holy grail is to get a thought process dataset"—capturing not just inputs and outputs, but the reasoning steps between them. Code provides exactly that because every line is testable.

THE STAKES: Poolside is simultaneously raising $2 billion at a $14 billion valuation—up from $3 billion just a year ago. Kant added: "This partnership ensures immediate access to next-generation silicon, enabling us to train multi-trillion parameter models with large-scale reinforcement learning."

WHY IT MATTERS: As AI agents get more capable, the amount of exploration they need decreases—just like an intern needs more guidance than a senior engineer. Poolside is building an "elastic AI workforce" that enterprises can scale up or down, orchestrating thousands of agents simultaneously.

Most AI companies are still building better chatbots. Poolside is building digital workers that learn by doing—and now they're securing the infrastructure to train them at industrial scale.

If they're right that intelligence scales with compute, West Texas might be where the first "AGI" gets built. One trial-and-error GPU cycle at a time.

ICYMI: here's last week's biggest stories at a glance...

💬 OpenAI Transforms ChatGPT Into an App Store

The Headline: At DevDay last week, OpenAI dropped a massive announcement: ChatGPT is no longer just a chatbot — it's becoming a full-blown platform where developers can build and sell apps. Think: iOS App Store, but for AI.

The Numbers: ChatGPT now has 800 million weekly active users. That's nearly the entire population of Europe using it every week.

What's New:

• Apps SDK: Developers can now create mini-apps that run inside ChatGPT with interactive interfaces
• AgentKit: Tools to build AI agents that can actually complete multi-step tasks (like scheduling appointments or writing code)
• Hardware Hints: Jony Ive (the guy who designed the iPhone) is reportedly working with OpenAI on physical AI devices

Why It Matters: This is OpenAI's play to become the next big tech platform. If they pull it off, ChatGPT could be where you start your day instead of your browser or phone home screen.

💰 OpenAI Strikes Massive Chip Deals Worth $100B+

The Scoop: OpenAI just inked two gigantic deals to secure computing power for its next-generation AI models. First, a partnership with Nvidia to build data centers using Nvidia chips. Then, days later, a similar deal with AMD worth tens of billions.

The Unusual Part: AMD isn't just selling chips — they're giving OpenAI the right to buy 160 million shares at one penny each (essentially 10% ownership of AMD if certain milestones hit). AMD's stock jumped 34% on the news.

The Controversy: Critics are calling these deals "circular" — companies investing in each other while also buying from each other. Some analysts are comparing it to the dot-com bubble. OpenAI says it's necessary to build the infrastructure needed for future AI breakthroughs.

Bottom Line: The AI infrastructure race is heating up, and OpenAI is betting big that they'll need absolutely massive computing power in the coming years.

🔒 Google's CodeMender AI Automatically Fixes Security Bugs

The Announcement: Google DeepMind unveiled CodeMender, an AI agent that doesn't just find security vulnerabilities in code — it automatically rewrites the code to fix them.

The Problem It Solves: Software vulnerabilities are notoriously hard and time-consuming for developers to find and patch. As AI gets better at discovering new bugs, humans alone won't be able to keep up.

Early Results: In the past six months, CodeMender has already pushed 72 security fixes to open source projects, including some with 4 million+ lines of code.

Why This Matters: If this works at scale, it could dramatically improve software security across the board. Instead of waiting months for developers to patch critical bugs, AI could fix them automatically within hours or days.

🖥️ Google Launches Gemini 2.5 Computer Use Model

What It Does: Google's new Gemini 2.5 Computer Use model can interact with user interfaces just like a human — clicking buttons, filling out forms, scrolling, typing, and navigating websites. Even the ones behind logins.

Use Cases: Imagine telling an AI "Go to this website, find all the California residents, and add them to my CRM, then schedule appointments for them" — and it just does it. That's what this enables.

Performance: Google claims it outperforms competing models on web and mobile control benchmarks, with lower latency.

Available Now: Developers can start building with it today via the Gemini API.

🏢 Anthropic Scores Biggest Enterprise Win Yet

The Deal: Anthropic (makers of Claude) just landed their largest enterprise deployment ever. Deloitte is rolling out Claude to 470,000 employees across 150 countries.

Why This Matters: When a Big Four consulting firm deploys AI to nearly half a million employees globally, it signals that AI has moved from "cool experiment" to "core business infrastructure." This is Anthropic winning major competitive battles against OpenAI and Microsoft in the lucrative enterprise market.

The Strategy: Enterprise customers care more about reliability and trust than being first to market. Anthropic's focus on safety and transparency is paying off in this space.

📊 Other Notable AI News from the Past 24 Hours

OpenAI's Broadcom Partnership

OpenAI announced they're co-developing custom AI chips with Broadcom to deploy 10 gigawatts of computing power. The chip shortage is real.

Ohio Wants to Ban AI Marriages

Yes, really. Ohio introduced legislation to prevent people from legally marrying AI systems. The future is weird.

AI in Healthcare Advances

New research shows AI can accurately screen for diabetic retinopathy (a leading cause of vision loss) before symptoms appear. This could increase screening access in rural areas.