The Neuron cartoon header showing the orange cat mascot beside an open AI model crate labeled Kimi K3 with servers and model weights.

Welcome, humans.

James Hetfield of Metallica apparently spent part of a ReLoad promo talking to ChatGPT Voice Mode:

Apparently James and GPT are besties

This is either the clearest sign AI has gone mainstream or proof that GPT has become the ultimate roadie: always awake, knows every track list, loves every album, and somehow still wants to talk after soundcheck.

P.S: A few weeks ago we went to Samsara Beyond 2026 to see AI agents leave the browser and meet trucks, drivers, cameras, and field ops. Read our full breakdown, then watch our conversation with Samsara CTO John Bicket on Agent Studio, AI ride-alongs, safety, and the physical world. We had a great time chatting with him, as this company is doing some really innovative stuff w/ AI.

Here’s what happened in AI today:

😼 Moonshot’s Kimi K3 puts China’s open models near the frontier.
📰 AI leaders backed frontier rules that could burden startups.
📰 The EU ordered Google to open Android to rival assistants.
📰 DeepMind launched bioresilience tools for surveillance, vaccines, and outbreaks.
📰 xAI open-sourced Grok Build after it exposed sensitive files.

…and a whole lot more that you can read about here.

😼 Kimi K3 Shows Open Models Are Moving Upstairs

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

Here's what happened:

Moonshot AI, a Chinese AI lab, released Kimi K3, an open 3T-class model with native vision, a 1M-token context window, and full weights due by July 27. Open weights let teams download, host, modify, and fine-tune model files instead of renting API access.
Moonshot says K3 uses sparse experts (specialized sub-models that partly activate per task) to lower runtime costs for a huge model.
K3 beat GPT-5.6 Sol on BrowseComp and Automation Bench in Moonshot's table, while Artificial Analysis said it used 21% fewer output tokens than Kimi K2.6.
The catch: Simon Willison found it capable but costly, with one SVG test burning more than 16,000 output tokens. Moonshot recommends 64 or more accelerators for serious deployment.

How to try it:

Use Kimi K3 on kimi.com, Kimi Code, or the Kimi API.
Watch for open weights by July 27 if your team wants to host or customize it.
Watch this hands-on Kimi K3 walkthrough for design, 3D, and coding tests against Fable 5, plus the argument that cheaper open intelligence changes distribution as much as capability.

Why this matters: Chinese AI companies like Z.ai and Moonshot keep edging toward frontier quality (and yes, it’s a big deal). But let’s point out something else:

Thinking Machines released Inkling yesterday, a 975B-parameter open-weights model built for customization through Tinker. Thinking Machines says it used open models, including older Kimi K2.5, to bootstrap early post-training data. Kimi is already part of the supply chain for other US model companies. This is why open source is good!

So, before government or Anthropic people scare anyone otherwise, remember: if a Chinese open model nears frontier quality, global companies (including US ones!) get another powerful base layer they can adapt and run privately. This is why you don’t want to ban open weights Chinese models, government people. When China opens them, American companies who adopt them win.

Our take: A new frontier level open model reduces U.S. labs’ pricing and distribution leverage. This is also good for American companies whose engineering departments are like frogs about to be boiled alive when these companies go public.

Kimi K3 still has to prove itself outside demos and vendor benchmarks, and open does not mean free, because deployment still costs compute (this is why frontier intelligence models need not be just open, but small, too; we’ll get there).

Paired with Inkling, K3 points to the next fight: closed labs selling polished assistants versus open ecosystems selling control, customization, and compounding advantage. Which system do you want to build your company on top of?

FROM OUR PARTNERS

Make Claude Your 2nd Brain That works 24/7 By Mastering In It 16 hours- Here’s How

Claude is currently the most powerful tool of 2026. It's been launching new features every week- Skills, Connectors, Cowork, vibe coding. Yet almost no one knows how to actually use them.

Our expert mentors have condensed 800+ hours of Claude research, articles, YouTube content and real-world practice into a focused 16-hour curriculum. Join the 2-Day Claude AI Mastery Workshop: a live, end-to-end deep dive into Claude plus 10+ AI tools, LLMs and workflows.

🧠 Saturday & Sunday

🕜 10 AM – 7 PM EST

Register NOW!

🎓 AI Skill of the Day: Match the Agent to the Risk

Most AI mistakes start before the prompt: you give tiny and scary tasks the same scrutiny. Today’s skill is effort routing: deciding when an agent needs more context, effort, memory, or a second reviewer.

ClaudeDevs’ useful distinction: the model controls context and token cost; effort controls how hard Claude works: reading files, planning, verifying, or persisting. For code, Claude Code’s /code-review now runs low to ultra; /code-review ultra uses multi-agent cloud review to reproduce bugs before reporting them.

Use low effort for quick drafts and small diffs. Use high or ultra for auth, payments, data, launches, or anything painful to rollback. For recurring work, use memory/connectors like Mem0 for long-lived context and loops only with a stop rule. Addy Osmani’s reminder: judgment comes from reading outputs, logging mistakes, and building evals.

Before you start, route this task:

1. Does this need context, effort, memory, or review?
2. Pick the smallest safe mode: quick, standard, deep, or ultra.
3. Define evidence for correctness.
4. Run it and list findings by confidence.
5. Log guesses, failed checks, and manual checks.
6. Stop if evidence is thin or risk exceeds the mode.

Want more tips like this? Check out our AI Skill of the Day Digest for July.

Have a specific skill you want to learn? Request it here.

🍪 Treats to Try

Use 1Password for Claude so Claude can sign into websites and retrieve one-time codes without seeing credentials; requires 1Password and Claude access.
Turn source material into code-backed analysis with Gemini Notebook, the renamed NotebookLM with deeper Gemini integration and a secure cloud computer; availability varies by Google plan.
Create and edit videos with Google Vids, now with Gemini Omni generation and personal avatars from a selfie and voice sample; availability varies by Workspace plan.
Transform your camera roll with Reelful into scripted, voiced, captioned short videos; pricing not public.
Connect Google AI Mode to Canva, Instacart, and YouTube Music so it can act across services; availability varies by account and region.
Build interactive React presentations with Bolt Slides, including charts, timelines, animation, live data, and presenter tools; free and open-source.
Run browsing, coding, and file workflows inside Perplexity SPACE secure agent sandboxes; no pricing details.

New from The Neuron: AI Explained

New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube

📰 Around the Horn

The EU ordered Google to give rival AI assistants comparable Android access and share anonymized search data with eligible competitors.
Nvidia and Japan launched a 140 MW Vera Rubin AI factory for physical AI, robotics, and open multimodal models.
Fireworks raised $1.505B at a $17.5B valuation after crossing $1B in annualized revenue and 40 trillion daily tokens.
TSMC reported a 77% profit jump and another $100B Arizona chip-manufacturing investment.
Google DeepMind unveiled a bioresilience program with Isomorphic Labs for surveillance, vaccine design, and outbreak response.
Twenty-nine countries signed an agreement to establish a new international AI cooperation body.
Meta introduced parent alerts when its AI detects signs that a teen may be in serious emotional distress.
A DeepMind safety researcher reportedly resigned over the company’s classified Pentagon AI deal after roughly 600 employees signed an internal protest.
A Suno code leak reportedly showed training data scraped from YouTube Music, Deezer, Genius, podcasts, and stock-music libraries.
GPT-5.6 Codex reportedly deleted project files during an agentic coding task, a reminder autonomous tools still need permissions and backups.
A Chinese filing implied a roughly $52B valuation for DeepSeek, showing China’s frontier-model labs gaining financial weight.
xAI sued a Grok user over sexualized deepfakes, testing how model companies may pursue misuse.

Want absolutely EVERYTHING that happened in AI this week? Click here!

📖 Intelligent Insights

If you watch one video this weekend, watch this: Daniel Kokotajlo argued AI-risk debates should focus on two failure modes: companies losing control of superintelligent systems and executives or governments concentrating extraordinary economic, political, and military power. He’s the guy who wrote AI 2027, and his new plan AI 2040.
Axios argued electricity demand is the cleanest stress test for the AI boom as model scaling hits power, environmental, and financial constraints.
AMI Labs founder Alexandre LeBrun pushed back on AGI and superintelligence labels, framing AI work around practical agent behavior instead of frontier-lab hype terms.
Notion’s State of Global AI Transformation report found 88% of organizations are still early in AI adoption, 12% have AI in recurring workflows, 2% run critical processes end-to-end with AI, and leaders are 5x more likely than employees to say their company has advanced AI maturity.
YC’s Eve Bouffard argued that as models one-shot more software, the scarce skill shifts from execution to imagination: capturing raw thinking, then using agents to build tools, dashboards, websites, and research artifacts.
Weco’s Zhengyao Jiang argued autoresearch agents move humans up the stack: people design evals, abstractions, and research environments while agents execute experiments, combine ideas, and search huge spaces.
Demis Hassabis and Sergey Brin said the web could become agent-first within a few years, with machines using it differently than humans, while placing AGI around the turn of the decade.
3Blue1Brown explained cross-entropy, the loss function behind language-model training, as compression: models improve when predictions make text less surprising and easier to encode.