Week of September 13, 2026
GPT-6 Astra lets non-programmers build 3D worlds in Blender, though even $200 Pro users burn through token limits. Plus why Apple's bet on on-device privacy boxed it out of the AI race.
This Week's Top Videos
GPT-6 Astra Does Everything. Here's What It's Actually Good At.
By AI For Humans
OpenAI's GPT-6 Astra makes computer use practical: the hosts report it operating Blender, GarageBand, games and cameras well enough that a non-programmer built a 3D Backrooms environment, work that used to take weeks. The catch is cost. Even $200 Pro users burn through token limits and banked resets, so budget usage before betting workflows on it.
- Astra's computer use has moved past fragile browser automation, where agents added items to an Instacart cart and then errored out. The hosts say it now runs desktop software like Blender and GarageBand and handles 3D modeling, textures, camera control and music creation.
- Comedian Duncan Trussell, who is not a programmer, used Astra to drive Blender and build a full 3D Backrooms environment. The hosts say this kind of scene used to take an artist weeks of manual work, which shows how far the bar for 3D creation has dropped.
- The hosts argue Astra makes much of the complex prompt engineering, agent pipelines and multi-step workflow scaffolding unnecessary. Builders should re-test whether their orchestration layers still add value or just add maintenance.
- Usage limits are the real constraint. Even the $200 Pro plan hits token caps quickly, and the hosts burned through multiple resets in one weekend. Plan your usage and cost before you build production workflows on Astra.
- Pro plan usage resets can be banked and used whenever you choose, but they carry expiration dates. Track them so they don't lapse unused. OpenAI's Tibo has also been handing out surprise resets to heavy users on X.
- Astra is not a clean win everywhere. The hosts note that some developers don't think it is as good at coding as rivals, so benchmark it against your current coding model before switching.
Why Has Apple Failed So Badly at AI?
By 20VC
Apple's AI problem is structural: it isn't a cloud company, and agents get better with more data that doesn't live on the phone. By tying its privacy brand to on-device models that are slower and dumber than the frontier, Apple boxed itself out. Builders: don't hardcode a values stance into one architecture.
- Apple's first disadvantage is cultural. Cloud is not in its DNA, and that gap now matters more than any single model or feature.
- Agents improve with access to more data, and most of a user's useful data lives off the device. Any agent strategy confined to the phone starts data-starved.
- Apple built its privacy and on-device story for competitive positioning, and that positioning now keeps it far from the frontier. A marketing differentiator can become a technical trap.
- Local on-device models are impressive but remain slower and less capable than frontier cloud models. Betting your product on local inference means accepting a capability gap.
- The privacy stance itself is good. The mistake was coupling it too tightly to on-device execution. Separate the principle you hold from the architecture you use to deliver it.
- Apple's privacy constraints make it hard to connect the data already on the phone across apps. That limits what an assistant can actually do, even locally.
I'm Obsessed With Local AI. Here's Why
By Greg Isenberg
The question to ask about local AI is not whether a model beats frontier cloud models but whether it is good enough for the job and whether running locally makes the product better. A 2B to 4B model can run on a phone or a 2021 laptop for private, offline, low-latency, or repetitive workflows, and that opens a wave of startup opportunities over the next 24 months.
- Decide where the intelligence should live. Use frontier cloud models for deep research and hard reasoning. Use local models for private files, sensitive customer data, offline field work, low latency, audio input, and internal workflows that run repeatedly.
- The local AI stack has four pieces. The model is the brain file (Gemma, Llama, Mistral, Qwen, Phi). The warehouse is Hugging Face. The runner is LM Studio or Ollama. The workflow is the product you build around all three, and that is where the business value sits.
- Read Hugging Face model cards with six questions. What is the model for, how big is it, what license does it use, what hardware do people run it on, which modalities and tool use does it support, and are quantized files available.
- Start non-technical users on LM Studio for a point-and-click chat experience. Builders should use Ollama, where one command such as 'ollama run gemma' gives you a local model with an API your apps can call.
- Use parameter count as a rough sizing guide. 2B to 4B models suit phones, edge devices, and fast small tasks. 12B is the middle ground. 26B to 31B is workstation territory. Run models on hardware you already own before spending $5K to $20K on a workstation.
- To move from a laptop demo to a shipped on-device app on iOS, Android, or web, study Google AI Edge and its LiteRT-LM runtime. Also learn the inference layers underneath: llama.cpp for general local inference and MLX for Apple silicon.
How to Talk About Price Without Losing the Deal
By The Futur
Say your price first or the client's number anchors you. Chris Do-style role play shows how to ask permission, state a range, and confirm fit before pitching anything. Price disqualifies fast, which saves everyone time. Walk away when six people must approve the deal, and always talk live, never by DM or email.
- State a big number first to set the price anchor. If the client opens with '100 bucks,' their anchor works against you, and dodging with 'it depends' makes you a 'slippery fish' who loses the call.
- Use a permission-based money script before pitching: 'Out of respect for your time, I need to tell you how much it costs to work with somebody like me. Is it okay if I tell you about this first?' Then give a range and ask if it fits their budget.
- Treat price as a disqualifier, not a closing line. If the budget doesn't fit, the conversation ends there, like a Ferrari lot shopper who knows to leave. Never open with 'Would you work with me for 1,000 euros?' because that forces one big decision too early.
- Move prospects through a series of small, easy yes/no decisions instead of one big ask. People stall when a decision feels like killing their other options.
- Qualify the decision maker early. When six people must approve a mid-sized project, walk away or quote roughly 10x, because committee buying will drive both sides crazy.
- Clients always know their budget. Surface it by offering tiered brackets (under 100k, 100k to 500k, over 1M), then narrow down. Do this live, because DMs and emails are 'a coward's way out' and clients can't hear your smile.
13 Designs that Inspired MIT Professor Zach Lieberman I Source Material
By Figma
Zach Lieberman says coding is a form of thinking, just like drawing, and the book that changed his life was John Maeda's Design by Numbers, a toy coding language. His other picks show that ideas start as napkin sketches, two frames are enough to animate anything, and knowing a tool exists can change your mind.
- John Maeda's Design by Numbers is the single most important book for Lieberman. It pairs code snippets with visuals and explains what coding feels like, not just how it works. That approach opened computation as an art medium for a generation of creative technologists.
- Treat coding the way artists treat drawing. Drawing Is Thinking argues sketching is cognition, and Lieberman applies that directly to writing code every day. Code is exploratory thinking, not just production.
- Everything starts as a napkin sketch. Lance Wyman, designer of the 1968 Mexico City Olympics logo, filled diaries with hand-drawn rough ideas, including icons for Mexico City subway stops, before they became iconic systems.
- Constraints are liberating. Masahiko Sato's Midnight Animation uses red, green, and white printing on black paper plus a blinking red/green DVD light to animate pages, proving you only need two frames to bring an image to life.
- Study original source material, not compressed copies. Lieberman prefers high-quality scans of abstract animation films over YouTube versions because compression hides the rich color and detail you need to actually learn from the work.
- Broad tool awareness expands what you can imagine building. Kevin Kelly's Cool Tools catalog is mostly irrelevant to Lieberman's own needs, yet he values it because simply knowing a tool exists reshapes your sense of what is possible.
When to use the New Agents vs A workflow with AI
By n8n
n8n now splits automation into three tiers: autonomous Agents with personalities, tools, and skills that live in Slack like a coworker, fully deterministic no-AI workflows, and semi-deterministic workflows where AI handles one defined step. Real production systems always end up needing deterministic logic, and the tiers interlock, so workflows become agent tools and agents run inside workflows.
- n8n's new Agents tab gives AI the most freedom on the platform. You define capabilities, skills, and a personality, and the agent decides when to use its tools for tasks without predefined steps.
- Example setup: a marketing assistant agent with Miro and Notion as tools and Anthropic's marketing skills loaded, deployed where the team already works, such as Slack.
- Constrain agent autonomy by limiting which tools it can access or by requiring human-in-the-loop approval on sensitive tools.
- Production systems always end up needing deterministic logic. Fully deterministic workflows run fast and cheap, behave identically every run, and are easy to audit visually.
- Use semi-deterministic workflows when AI should be a defined step with a specific job, such as classification or filling out data, rather than an autonomous coworker.
- Stop building chat assistants out of workflows. Instead, expose any workflow with the right trigger as an agent tool, or drop the same agent onto the canvas to hand off tasks or power webhook-based embedded support bots.