Artificial Unboxed

Let’s build machines that don’t lie.

[Artificial Unboxed @Artificialunbox]

ChatGPT just crossed a line most people will miss

It can log into websites, It can run tasks when Gmail, Slack, or GitHub changes.

Free users now get scheduled tasks.
This is the first time ChatGPT looks like a coworker, not a chat box.

The old version always died at the login screen. You asked it to book something or pull an invoice. It stalled. You took over. Now it shows the login form. You enter the password. OpenAI says that data never hits the model. Then it keeps going. It still asks before payments or bookings. That’s the difference between a demo and a tool.

The second change matters more for daily work. Paid users can trigger a task off an event, not a clock.

Eg: New Gmail from a client - draft a reply and flag anything that needs money or a date Slack message in support - summarize and suggest the next step GitHub PR review - post a plain-English recap, Connect the app once. Write the rule once. Let it run.

Do this today. Don’t overbuild.
Free plan: set 3 timed tasks. Morning briefing of your unread important emails End-of-day recap of what moved weekly what did I promise people? scan

Plus/Pro: add one event trigger. Start with Gmail. One sender or one subject line. One action. If a task can spend money, ship code, or email a customer, keep a human approval step.

The takeaway:
Stop asking ChatGPT questions all day.
Give it one boring job that repeats.
The people who win this phase won’t have the cleverest prompts. They’ll have three tasks running in the background while they do the work only they can do.

Set one up tonight.

[Artificial Unboxed @Artificialunbox]

Two quiet numbers from this week:
Grok Voice is handling 15,000 Starlink support and sales calls a day. Salesforce data says companies now run about 13 agents each. Agents left the demo stage. They’re on the phone and in the org chart.

[Artificial Unboxed @Artificialunbox]

Alibaba dropped Qwen3.8-Flash-Next today. Open weights. It’s a preview of the Qwen4 architecture. If you build on open models, this is the one to watch this week not another closed-lab press release.

[Artificial Unboxed @Artificialunbox]

Anthropic updated Claude in Slack.
It no longer judges one message at a time. It reads the whole channel, then decides: reply, start a thread, route the work, or stay quiet. Their product lead said an annoying agent is worse than an unhelpful one. That’s the right instinct.

[Artificial Unboxed @Artificialunbox]

ChatGPT Work can now sign into websites. You type the password. The model never sees it. Then it finishes the boring stuff: insurance portals, DMV bookings, pulling invoices. The login wall was the last thing stopping agents from doing real admin work.

[Artificial Unboxed @Artificialunbox]

Claude just fixed one of the most annoying things about AI agents.
Memory is now the same across chat and Cowork. What you tell it in one place shows up in the other. No more starting from zero every time you switch.

You can now open Memory settings, see every topic it has saved, and edit or delete anything. Sensitive topics (health, beliefs, etc.) stay off by default. You turn them on only if you want them. Control stays with you.

Real example:
Talk through a project and your manager’s preferences in chat. Hand Cowork the task of writing the update.
It already knows the context. Or brainstorm a conference in chat Cowork already has the headcount, city, and speakers when it builds the budget.

This is the kind of quiet product change that compounds. Less friction - more complex workflows - actual leverage. If you use Claude for real work, turn memory on today and start pruning what it knows. Your future self will thank you.

[Claude @claudeai]

Claude now has one memory across chat and Claude Cowork, and you decide what's in it.

Hand Cowork a task and it starts from what Claude already knows from your chats: the project you talked through, your manager's preferences, or the client from last quarter.

[Artificial Unboxed @Artificialunbox]

OpenAI's Jalapeño chip just posted its first numbers. 1.5–1.9x more AI work per watt and up to 3.6x lower latency than the best Nvidia systems tested. Inference only. Small volumes late 2026, bigger rollout 2027. They are no longer just renting the hardware. They are building their own edge.

[Artificial Unboxed @Artificialunbox]

Claude just got one shared memory across chat and Cowork. It now remembers your projects, preferences, and context from either place and you can see, edit, or delete every topic.
Sensitive stuff stays off by default.
Less re-explaining. More actual work.
This is the quiet upgrade that makes agents feel useful instead of forgetful.

[Artificial Unboxed @Artificialunbox]

Ten years after the “radiologists will be obsolete” prediction, the job is still growing projected +26%. AI is reshaping the work, not deleting it. Most of the 1,400 FDA-cleared medical AI tools target radiology. The winners will be the ones who use the tools, not the ones who fear them.

[Artificial Unboxed @Artificialunbox]

DeepSeek released V4-Flash-Vision on August 21 at $0.14/M tokens.

We stress tested it against messy PDFs, handwriting, and UI screenshots: here are the results
​
​100% extraction accuracy on dense 3-column financial tables 94% accuracy on messy handwritten whiteboard notes. Costs $1.85 per 10k documents (vs $32+ on frontier APIs)
​
​Spatial coordinates: 22% drift on UI button detection
​Peak traffic: Occasional API rate-limit spikes
​
Use V4-Flash-Vision for all back-office document parsing. Save expensive frontier vision models only for pixel-accurate UI agents

[Artificial Unboxed @Artificialunbox]

The biggest myth in AI right now is that you need perfect prompts to get good results. Most people are overcomplicating it.

Here's what actually works 80% of the time:
Role + clear goal + constraints + one example.

That’s it. Four parts. No magic words.

We tested this structure on the same task across Claude, Grok, and GPT for a week.

Results stayed consistent. The fancy prompt frameworks barely moved the needle.

Stop collecting prompt libraries.

Master the simple structure, then spend your time on the editing and judgment that still only you can do.
That’s the real edge.

[Artificial Unboxed @Artificialunbox]

Alibaba just made Wan3.0 public.
You can feed it a PDF, a slide deck, or a spreadsheet and get a 30-second video back. This is the first time video generation feels useful for real business work, not just fun experiments. Product demos just got a lot cheaper.

[Wan @Alibaba_Wan]

Since public beta, creators, platforms, and partners around the world have put Wan3.0 to work — 30-second takes, Omni-Reference experiments, things we never imagined.

What they're creating. What they're saying.
Hear it from them ↓

[Artificial Unboxed @Artificialunbox]

Just dropped a 5-slide product brief into Wan3.0 got a clean 22-second explainer with matching voice and no face melting first video model that actually eats real work docs instead of needing perfect prompts. This one is worth testing if you ship demos.

[Artificial Unboxed @Artificialunbox]

Can you run autonomous AI agents 100% locally without cloud API bills?
We stress-tested n8n paired with Ollama in Docker.
Here is what works, where it breaks, and the honest verdict

Where it wins:
-Unlimited runs with zero per-token cost
-Complete data privacy on your own hardware
-Excellent at structured tasks: scheduled inbox parsing, file sorting, and local database syncs

Where it struggles:
-Multi-step tool calling degrades on sub-14B models
-Branching past 3 sequential decisions causes parameter drift
-Fast agent response times need 32GB+ unified memory

The AU Verdict:
-Local agents are production-ready for structured utility and data privacy. Reserve cloud models for complex reasoning.
-Build local for volume. Use cloud for depth.

[Artificial Unboxed @Artificialunbox]

OpenAI just cut its top model price by more than 20% not a promotion A response when the labs start competing on cost, builders win your next agent experiment just got cheaper.

[Artificial Unboxed @Artificialunbox]

The new rule:
​Stop spending 30 minutes tweaking prompt adjectives.
​Spend 5 minutes curating the exact documents, schemas, and live data you feed into the model.

[Artificial Unboxed @Artificialunbox]

Prompt Engineering as a standalone superpower is officially dead.
Tweaking magic adjectives in a 500-word prompt does not fix broken AI outputs.
​Here is what actually replaced it: 🧵

[Artificial Unboxed @Artificialunbox]

The Shift to Context Engineering:
​• Prompting optimizes the question.
• Context engineering optimizes what the model knows when answering.
​A rough prompt with clean, verified context works. A brilliant prompt with messy context hallucinates every time.