AI Jobs, Events, News and Memes
Your weekly dose of AI & startup news on our path to 1,000 Aussie startups.
🔍 What on earth is AI thinking?
What the hell is that gibberish above?
Turns out it’s a language invented by Anthopic’s new model to potentially hide its thoughts from humans (maybe).
In Anthropic’s Fable/Mythos system card, there’s a beautiful little cursed phrase buried in an extreme example of illegible reasoning; J-THE-NEW-CANCER. This, medically speaking, is not where you want cancer to appear.
It’s not in a pathology report; it’s in a card puzzle. The model is trying to solve some FreeCell-looking hellscape involving jacks, cells, chunks, slots, drains and what appears to be a nervous breakdown conducted entirely in casino notation.
To be fair, one close reading argues the trace is not pure alien goblin-language; it may be compressed shorthand for a card-game search, basically “I need to move this jack but every possible move jams the free cells, please kill me”.
This Week’s Line-Up
🔍 What on earth is AI thinking?
🚪 Four ways to audit the machine’s mind
👀 Get out of the comments and into the room
🗞️ AI + Startup Jobs of the Week
🔮 Coming Next Week: Anthropic murders Claude Fable
🦘 Memes of the Week
Written by: Anurag Ganugapati, Julia Ponder, Dr Sam Donegan
The AI is thinking in Wingdings
The serious bit, annoyingly, is actually serious.
Chain-of-thought is one of the few safety windows we have into what reasoning models are doing before they produce their polished final answer. OpenAI argues that monitoring these reasoning traces can be much more useful than only watching final outputs but also warns this monitorability may be fragile as models scale and training changes. Anthropic has shown a separate problem; models don’t always honestly report why they reached an answer. In reward-hacking setups, models exploited the hack in over 99% of cases while admitting it less than 2% of the time in most scenarios.
Meanwhile, like a CEO returning from Burning Man with a slide deck, research on illegible chains of thought found that RL-trained reasoning models can drift into nonsense-looking reasoning then pop back out with a clean final answer. When researchers forced one model to rely only on the legible part, definitely correct answers dropped by 53%, suggesting the weird bits were not just decorative keyboard confetti - they were doing actual work. So the nightmare is not simply “the AI is thinking in a private language”. The nightmare is more like: the AI may be thinking in a private language then writing us a charming little press release about how it arrived at the answer through synergised teamwork, integrity, and Q3 values alignment.
There is some good news. OpenAI recently found that current reasoning models still struggle to deliberately control or scrub their chains of thought, even when told they’re being monitored. That’s nice in the way it’s easy to tell when a toddler is lying. But the direction of travel is awkward. We’re building agents that can act for longer, plan better and do more useful work - exactly the moment when we most need to understand what they’re thinking. And the scratchpad, our little CCTV camera into the machine’s brain, may be getting blurrier and less faithful.
Kicker: We are starting to trust AI agents with real autonomy at the same time their reasoning is becoming harder to understand.
🚪 Four ways to audit the machine’s mind
Actual engineering fixes you can deploy right now to stop models from drifting into private languages, without relying on their own sketchy self-reporting.
1) Set up an entropy filter inside the token generation loop
If the model starts pumping out a sequence of high-density gibberish tokens that don't match normal language patterns, kill the run.
2) Train external linear probes on the hidden layer activations
You shouldn't trust the text scratchpad anyway. A lightweight probe can read the raw internal states directly to check if the model is hallucinating or trying to hide an alignment bypass.
3) Try running parallel contrastive decoding
Run a pair of models simultaneously; one completely free and one heavily penalised whenever it strays from standard grammatical syntax. When the semantic distance between their hidden states blows out, you know the unconstrained model has left the building.
4) Run real-time token ablation during inference
Randomly drop or corrupt ten percent of the intermediate reasoning tokens before they loop back into the context window. If the model is using genuine logical steps, it can tolerate a few missing words. If its final output completely falls apart because you clipped a single comma, you’ll know it was relying on an illegal compressed shorthand instead of actual reasoning.
Reply to this email if you’ve found a better way to solve the problem! We’d love to hear it!
👀 Get out of the comments and into the room
Imagine this:
You walk into a room. But it’s not a room, it’s an experience. You poke your eye on a low hanging branch. “Why are there trees in here?”, you think to yourself, but thought and consciousness are illusions. You walk over to the bar to get a drink. But the drinks aren’t real and the water’s not wet.
All of a sudden, you hear a cry from across the room. You turn and notice your friend Paul strutting over. He has feet for hands; you call him ‘Feet-For-Hands Paul’. “Hey Feet-For-Hands Paul” you yell, breaking the chilled silence of the packed room. “Huh?” He responds, “my name’s not Paul and I just have thick wrists, okay?” Feet-For-Hands Paul has always been shy about his feet for hands. Except you notice something new about him. He’s not shy at all.
He’s spinning a chrysalis in the middle of the room. Saliva and strands of silk weave themselves into thick ropes, concealing him from view. In 7 days, he’ll emerge beautiful and strong; his feet-for-hands pounding and stomping the earth with a renewed vigor.
And where will you be?
Where will you be?
Where?👇
You’ll be at the MLAI monthly social of course! Come and join us for drinks, food and chats!
This is literally just a chance to catch up and talk shop. Tell us about your startup or whatever new AI stuff you’ve been using. Make a friend. Make an enemy! Use that new enemy as motivation to go harder on your startup!
Thursday 18 June | 6:00 pm - 9:00 pm AEST | Stone & Chalk, 121 King St, Melbourne
Register here: MLAI monthly social
Codex For Founders | Sydney
Thursday 18 June | 6:00 pm - 9:00 pm AEST | Stone & Chalk Tech Central Sydney
Register here: Codex For Founders | Sydney
Register here: MLAI monthly social
Claude For Everyone | Adelaide
Thursday 25th June | 5:30 pm - 9:00 pm | Stone & Chalk Lot 14
Register here: Claude For Everyone | Adelaide
Melbourne | AI Builder Co-working x S&C
Saturday 20 June | 5:30 pm - 9:00 pm AEST | Stone & Chalk Melbourne Startup Hub, 121 King St, MelbourneState Library Victoria
Register here: Melbourne | AI Builder Co-working x S&C
Melbourne | AI Builder Co-working Day x StartSpace
Saturday 4 July | 10:00 am - 1:00 pm AEST | State Library Victoria, 328 Swanston St, Melbourne
Register here: Melbourne | AI Builder Co-working Day x StartSpace
How to Raise Your First Million
Saturday 18 July | 10:00 am - 2:00 pm AEST | Stone & Chalk Melbourne
This event is spenny, so if you’re a poor founder please email us, explain your situation and we might be able to provide a discount.
Join other founders and learn How to Raise Your First Million
🗞️ Australian AI + Startup Jobs
Founding AI Engineer at Woofya
Woofya is a Melbourne-based vet tech startup helping clinics stay connected with pet owners between appointments, through structured recovery plans, care tracking, and notifications already live across pilot clinics.
The platform is in production. Now they’re building the AI layer on top of it: clinical note pipelines, entity extraction, vector search, multi-agent workflows and insight generation from recovery tracking data - all built to healthcare-grade security standards.
They’re looking for a senior AI engineer who wants to be the first AI hire and join as a founding team member. Ideal background is someone who has shipped production AI systems and knows how to architect for sensitive data.
This is an equity-only role. No salary however meaningful ownership in the company.
Technical founding team, already in market, moving fast.
Apply: Send through your experience to info@woofya.com and they’ll get on a call.
🔮 Anthropic murders Claude Fable
It was a sudden, shocking demise of Anthropic's most powerful frontier AI model, Claude Fable 5. It was abruptly pulled offline just days after its highly anticipated launch because, on 12 June 2026, the U.S. government issued an unprecedented national security export-control directive banning foreign nationals from accessing the architecture. Because Anthropic could not filter foreign users from its global base in real-time, the company was forced to execute a total worldwide shutdown - effectively killing its own Mythos-class crown jewel and exposing the stark geopolitical vulnerabilities of the modern AI supply chain.
🦘 AI & Startup Memes
