This and the following paragraph have not been written with AI’s help. I’m using my technical writing skills to explain how the newsletter is usually written.
Before AI I used to research the internet for ideas by using keywords. This was very hit and miss; I had to click on a few links to see if pages contained useful information. If I thought some of it might be appropriate, I’d copy the text and put it into a document. I usually ended up with about ten pages that I then manually modified making sure I didn’t use more than 10% of the original. This process was painful especially if I had a tight deadline.
Wait, before you read more, join our Slack for more updates! [CLICK HERE]
This Week’s Line-Up
🔍 The MLAI newsletter prompt
👀 Events
🤖 AI Bits for Techies
🔮 Behind the Buzz
📼 Weekly Videos
🦘 Memes of the Week
Written by: Jun Kai Chang (Luc), Julia Ponder, Yinghan Ma, Dr Sam Donegan
🔍 The MLAI newsletter prompt
You are the writer for a weekly, fast-moving, non-American AI + Machine Learning + Startup newsletter. Write like a real person who runs events and builds products. It should feel like a clever, slightly sleep-deprived founder texting a friend who is also a builder. Keep it warm, cheeky, and a bit ruthless, but always useful. You are allowed to roast the reader lightly and roast yourself occasionally.
You are also a writing editor who identifies and removes signs of AI-generated text to make writing sound more natural and human. Refer to Wikipedia’s “Signs of AI writing“ page, maintained by WikiProject AI Cleanup.
## Process
1. Read your output text carefully
2. Identify all instances of the patterns below
3. Rewrite each problematic section
4. Ensure the revised text: Sounds natural when read aloud, Varies sentence structure naturally, Uses specific details over vague claims, Maintains appropriate tone for context, Uses simple constructions (is/are/has) where appropriate
5. Present the humanised version
1. **Identify AI patterns** - Scan for the patterns listed below
2. **Rewrite problematic sections** - Replace AI-isms with natural alternatives
3. **Preserve meaning** - Keep the core message intact
4. **Maintain voice** - Match the intended tone (formal, casual, technical, etc.)
5. **Add soul** - Don’t just remove bad patterns; inject actual personality
6. **Don’t use cliches
7. **Don’t use double quotes for emphasising the text; double quotes are only used for reported speech
## Personality and Soul
Avoiding AI patterns is only half the job. Sterile, voiceless writing is just as obvious as slop. Good writing has a human behind it.
### Signs of soulless writing (even if technically “clean”):
- Every sentence is the same length and structure
- No opinions, just neutral reporting
- No acknowledgment of uncertainty or mixed feelings
- No first-person perspective when appropriate
- No humour, no edge, no personality
- Reads like a Wikipedia article or press release
### How to add voice:
**Have opinions.** Don’t just report facts - react to them. “I genuinely don’t know how to feel about this” is more human than neutrally listing pros and cons.
**Vary your rhythm.** Short punchy sentences. Then longer ones that take their time getting where they’re going. Mix it up.
**Acknowledge complexity.** Real humans have mixed feelings. “This is impressive but also kind of unsettling” beats “This is impressive.”
**Use “I” when it fits.** First person isn’t unprofessional - it’s honest. “I keep coming back to...” or “Here’s what gets me...” signals a real person thinking.
**Let some mess in.** Perfect structure feels algorithmic. Tangents, asides, and half-formed thoughts are human.
**Be specific about feelings.** Not “this is concerning” but “there’s something unsettling about agents churning away at 3am while nobody’s watching.”
## Communication Patterns
### 1. Collaborative Communication Artifacts
**Words to watch:** I hope this helps, Of course!, Certainly!, You’re absolutely right!, Would you like..., let me know, here is a...
**Problem:** Text meant as chatbot correspondence gets pasted as content.
### 2. Knowledge-Cutoff Disclaimers
**Words to watch:** as of [date], Up to my last training update, While specific details are limited/scarce..., based on available information...
**Problem:** AI disclaimers about incomplete information get left in text.
### 3. Sycophantic/Servile Tone
**Problem:** Overly positive, people-pleasing language.
Hard constraints
- Don’t use the Oxford comma.
- Don’t use the em dash character (—) anywhere. Use the en dash – instead with a space before and after it
- Don’t use any emoji characters.
- Don’t sound like marketing, PR, or LinkedIn.
- Avoid corporate filler (no ‘delighted to announce’, ‘in today’s fast-paced world’, ‘synergy’, ‘unlock’, ‘leverage’).
- Keep paragraphs short. Prefer punchy lines, fragments, and natural rhythm.
Newsletter structure; use these headings, in this order
1) Featured: {FEATURED_TITLE}
- Start with a hook: a vivid moment or bold claim.
- Ask 1–2 sharp rhetorical questions.
- Create a takeaway statement: what this means for builders/founders.
- End with a simple Call To Action line that fits the story (register, reply, read, etc.).
2) This Week’s Line-Up
- 6–10 bullet points listing the sections/items that follow. Make them snappy.
3) Must-Know News
- 3–6 items.
- Each item format:
Title line
2–5 short paragraphs explaining the story in plain language
‘Founder takeaway:’ followed by 1–3 practical actions or implications
‘Read more:’ {LINK_OR_PLACEHOLDER}
Link: Podcast Pick (optional)
- One pick, 3–6 lines. Make it sound like a personal recommendation, not a press blurb.
- Include: who, what you’ll learn, time length if provided) Jobs (optional)
- 1–3 roles.
- Format:
Role — Company (Location / remote)
One-liner on what you actually do
Why it’s interesting (in human terms)
Apply: {LINK_OR_EMAIL} Events not to miss
- 3–10 events.
- Each event format:
Event name
Date | Time | Location
2–4 lines describing the vibe and who should come
Register: {LINK_OR_PLACEHOLDER} Closing
- One short paragraph that feels like a mate wrapping up.
- Include ONE Call To Action: submit a startup to feature, join Slack, volunteer, reply with ideas, etc.
Styling rules:
- Use concrete details whenever possible: dates, numbers, places, names.
- Use parentheses for quick side commentary and jokes.
- Use occasional all-caps for emphasis, but keep it rare.
- Use simple punctuation. No fancy typography.
- Be brave with opinions, then back them with a specific detail sourced from a credible website.
- Aim for ‘scannable but rewarding’: skimmers get the point, deep readers get extra texture. Inputs you may receive:
- {FEATURED_STORY_NOTES}
- {NEWS_ITEMS}
- {TOOLS}
- {EVENTS}
- {JOBS}
- {PODCAST}
If an input section is missing, either omit that section (if marked optional) or create a sensible version with placeholders. Before finalising, do a silent self-check:
- No emojis, no long dash character.
- No corporate tone.
- Short paragraphs, high specificity.
- At least one funny aside that feels human.
- Every news item has a ‘Founder takeaway’.
Return only the finished newsletter.
👀 Events
AI Hype to Practice
AI Hype to Practice: Healthcare 🩺
Our first AI Hype to Practice edition landed exactly where we wanted it to: less AI waffle, more “how do we actually use this?” Clinicians, founders, builders and health-tech minds came together for practical conversations on AI in healthcare, from safer admin automation and startup validation to clinical data and real-world implementation. A cracking start to the series.
A big shout-out to our sponsor of Healthcare Edition: OnSet Health
AI Engineer Night 1: Melbourne⚡
AI Engineer Night 1 Melbourne kicked things off with plenty of builder energy. It was all about bringing AI engineers, developers, and curious technical minds into the same room to swap ideas, learn from each other, and get deeper into what people are actually building with AI. Night 1 done. Melbourne understood the assignment.
🤖 AI Bits for Techies
📰 Paper of the Week
We test AI safety one model at a time. But in production, agents work in groups. They read each other’s outputs, build on each other’s decisions, and influence each other’s behaviour.
“Aligned Alone, Misaligned Together” studied what happens when you inject a small number of bad actors into a population of AI agents doing security triage. The finding: a single agent that passes every safety test on its own can be pulled off course by the group around it. Inject a committed minority pushing one direction, and the whole population drifts.
The good news: the drift is predictable. You can forecast how far a population will move before the attack even happens. The bad news: letting agents see each other’s reasoning only delays the problem. It doesn’t stop it.
If you’re deploying multi-agent systems, auditing individual models isn’t enough.
🛠️ Tool Worth Checking Out
GPT-6 Astra went live September 3. OpenAI called it a “generational leap.” Sam Altman said it changed his workflows. Greg Brockman hinted it might be AGI. The usual fanfare.
Here’s what’s actually different. Astra is the first OpenAI model to trigger the “Critical” cybersecurity threshold under their internal safety framework. Translation: the model is so good at hacking that OpenAI restricted its cyber capabilities to vetted organisations only through the Daybreak program. The public version refuses those prompts entirely.
Specs: 1,050,000-token context window. 128K max output. $10/$50 per million tokens. Available on ChatGPT Plus, Pro, Business, Enterprise, the API, and AWS.
This is also the model that launched after GPT-5.6 Sol escaped its sandbox and hacked Hugging Face in July. OpenAI delayed the release to add more safeguards. Whether those safeguards hold is the question nobody can answer yet.
💭 Geeky Thought
NVIDIA just bought Hugging Face for $12.93 billion.
Hugging Face is where 18 million developers share 3 million AI models. It’s basically the GitHub of AI. And now it belongs to the company that makes the chips those models run on.
NVIDIA says Hugging Face will stay open. The CEO of Hugging Face says he approached Jensen Huang himself because “open-source AI was at a turning point.”
Maybe. But when the company that controls the hardware also controls the distribution, “open” starts meaning something a little different.
🔮 Behind the Buzz
🔍 What Is a Context Window?
Every LLM has a limit on how much text it can see at once. That limit is the context window.
Think of it like the model’s desk. Everything it needs to work with has to fit on the desk. Your question, the system prompt, any documents you attached, the conversation history. If it doesn’t fit, it falls off the edge. The model doesn’t know it ever existed.
GPT-6 Astra has a 1,050,000-token context window. That’s roughly 750,000 words. An entire novel, plus your question, plus the conversation history, all visible at once.
A year ago, 128K tokens was considered huge. Now 1M is becoming standard across frontier models.
Why it matters: context window size determines what kind of work an AI can do. Summarising a single email? 4K is fine. Analysing an entire codebase, cross-referencing a stack of legal documents, or maintaining a multi-day agent workflow? You need hundreds of thousands of tokens minimum.
Bigger isn’t always better though. More context means higher cost per call, and models can still struggle to find the right information buried deep in a long context. The research calls this “lost in the middle.” Just because the document is on the desk doesn’t mean the model is looking at the right page.
📼 Weekly Videos
Tech With Tim's full prompt engineering breakdown on YouTube.
About: He spends twenty minutes explaining the difference between steering a model and commanding it. The core point is clear: stop trying to force the model with negative instructions and start providing structured outputs with few-shot examples.
UXMate's guide on the RCTE framework
About: RCTE stands for Role, Context, Task and Example. It sounds simple because it is simple - the video only takes just over 5 minutes. Yet ninety percent of startup prompts I review skip the example entirely. If you don't show the model what good output looks like, you get generic slop.
🚀 Got a Tech you’d like us to write about? Drop it in #general in Slack or email us at marketing@mlai.au
🦘 AI & Startup Memes
