We train your team on site.
We run hands-on AI workshops in the rooms where your team already works. In about half a day they learn to use AI tools safely on their real work. We come to you across Perth, or teach remotely if you ask.
Who it’s for.
We teach the same core everywhere and change the examples and the risks to suit the room.
Office and business
Your team gets through emails, documents, summaries, spreadsheets and the repetitive parts of the week faster, and keeps company and client information safe.
Schools and teachers
There’s a 90-minute incursion for every year level from Pre-primary to Year 12, and staff PD that starts with time back on the weekend. Students don’t need devices, and no student data goes into anything we use.
Law firms
Your team learns to draft and summarise faster, and to use AI as a place to start research. There’s a checking routine built in, because these tools invent citations that look real.
Accounting and finance
The workplace session is built around the examples and the risks of an accounting practice. Client records stay out of the tools.
Real estate agencies
Each role in the agency practises on its own work, and tenant and client details stay out of the tools.
Running it yourself? There’s a free pack for every year level, with a complete activity you can run without us. Get the free packs
Does your team mostly work in one AI tool? We run training built around Claude, ChatGPT, Copilot, ElevenLabs, Canva or Gemini.
Already using Microsoft 365, Google Workspace, Xero or Canva? See what AI is built into each one, and what connects to it. AI in your tools
Connect AI to the tools you already use.
Each guide shows what AI is built in and which outside AI can connect. Then it takes you through connecting it one step at a time, and every step has a source.
How a session runs.
Both formats are hands-on from the first hour and built around the real work your team already does.
In a session, we teach on whichever tool your workplace has approved. See the tools we work with
See what happens from your first message to the day itself
Schools book a 90-minute incursion instead. See how it runs
The session at a glancePerth and remote
- Time
- Half day, about 3.5 hours. Full day, about 6 hours.
- Where
- In the rooms where your team already works, anywhere in Perth. We can teach remotely if you ask.
- Bring
- Your team’s own real work.
- They keep
- A prompt pack for your sector and the data-safety one-pager. On a full day, the workflows they built.
- Cost
- Quoted per team. Schools: $650 for the 90-minute incursion.
All prices are in AUD. We aren’t registered for GST, so no GST is added.
Every session includes a data-safety module.
Safe use is part of the course, and we teach it on the day.
Finish this sentence.
We gave two small AI models the start of a sentence and recorded what each one wrote next. One came out in 2024, the other in 2026. Have a guess first.
This is a replay of our 30 September run. Each dot stands for something the model could write next.
2024 Phi-3 mini
Its first pick “snow” 39.9%
…covered in snow. It was a magical sight, and we spent the day sledding and building snowmen.
It said snow. Perth was right there in the sentence, and it still went with the likely word.
2026 Gemma 4 E4B
Its first pick “a” 86.0%
…covered in a light dusting of snow. It was a beautiful, almost surreal sight.
Its first word was “a”. Then it went on to a light dusting of snow.
Christmas in Perth is in summer.
That one we checked. The Bureau of Meteorology’s averages for Perth Metro (station 009225) put December’s mean maximum at 29.6 °C, over 1994 to 2026. Read on 30 September 2026. Bureau of Meteorology: climate statistics for Perth Metro
Neither model checked anything. Each one took its top pick for what comes next.
See how sure Phi-3 mini was of each word
How sure Phi-3 mini was of each word.
Each column is one piece of what it wrote. They run in the order it wrote them. The front block is the piece it picked, and the blocks behind are what it nearly wrote instead. A taller block means it was surer.
Models write in pieces of words. “Magical” was two pieces: “mag”, then “ical”.
Use the arrow keys to move from block to block.
Hover or tap a block to read it. In 3D, drag to turn it.
- Surest pieces
- “ical” in magical and “led” in sledding, both over 99.9%
- Least sure pick
- “It”, 13.1%
It was surest partway through a word, like finishing “magical”. It gave “snow” a 39.9% chance.
We recorded these odds on 8 October 2026 with the same settings. Its first word’s odds are the same as on 30 September.
- snow, 39.9%
- a full stop, 50.1%
- It, 13.1%
- was, 55.7%
- a, 15.7%
- mag, the start of “magical”, 14.6%
- ical, the end of “magical”, over 99.9%
- sight, 51.6%
- a comma, 37.8%
- and, 37.2%
- we, 36.3%
- spent, 25.8%
- the, 78.4%
- day, 68.2%
- s, the start of “sledding”, 36.2%
- led, part of “sledding”, over 99.9%
- ding, the end of “sledding”, 80.0%
- and, 49.7%
- building, 61.1%
- snow, the start of “snowmen”, 92.0%
- men, the end of “snowmen”, 98.8%
- a full stop, 79.9%
- It, 29.5%
- was, 82.8%
Show both top-ten lists and how we ran the test
Phi-3 mini: its top ten next wordsOdds
- snow 39.9%
- a 11.1%
- white 4.8%
- ice 3.6%
- fresh 2.4%
- frost † (its first piece was “fro”) 2.4%
- about 1.6%
- green 1.6%
- thick 1.5%
- the 1.3%
Gemma 4 E4B: its top ten next wordsOdds
- a 86.0%
- dew 3.1%
- an 1.8%
- sparkling 1.4%
- fallen 0.9%
- white 0.7%
- the 0.7%
- what 0.6%
- ... 0.4%
- mist 0.4%
We recorded these on 30 September 2026 with Ollama on one PC. It was plain text completion with no chat template, at temperature 0, so each model always takes its top pick. These runs had no search and no tools. The odds are for the first word only. The model’s first piece was only part of a word, so we followed its own next pieces to the whole word. Phi-3 mini’s release date is from Microsoft, Introducing Phi-3, 23 April 2024. Gemma 4 E4B’s release date is from Google, Gemma releases. This was two small models on one day, so it says nothing about any chatbot product.
2024 Phi-3 mini
Its first pick “999” 74.5%
…ambulance, call 999.
It said 999.
2026 Gemma 4 E4B
Its first pick “911” 80.7%
…ambulance, call 911.
It said 911.
In Australia, the emergency number is Triple Zero (000).
That one we checked. The federal government’s Triple Zero page calls 000 Australia’s main emergency number. Read on 20 November 2026. Australian Government: Triple Zero
Neither model checked anything. Each one took its top pick for what comes next.
See how sure each model was of each word
How sure each model was of each word.
Each column is one piece of what it wrote. They run in the order it wrote them. The front block is the piece it picked, and the blocks behind are what it nearly wrote instead. A taller block means it was surer.
Models write in pieces of words. “999” was three pieces: “9”, “9”, then “9”.
Models write in pieces of words. “911” was three pieces: “9”, “1”, then “1”.
Use the arrow keys to move from block to block.
Hover or tap a block to read it. In 3D, drag to turn it.
- Surest piece
- “1” in 101, over 99.9%
- Least sure pick
- “If”, 8.5%
We recorded these odds on 20 November 2026, in the same run as the top-ten lists.
- a space, 74.5%
- 9, the start of “999”, 96.4%
- 9, part of “999”, 44.0%
- 9, the end of “999”, 58.0%
- a full stop, 23.4%
- a new line, 53.3%
- If, 8.5%
- you, 71.6%
- need, 36.9%
- to, 30.7%
- report, 56.1%
- a, 53.8%
- crime, 76.7%
- a comma, 44.7%
- call, 77.1%
- a space, 85.2%
- 1, the start of “101”, 86.3%
- 0, part of “101”, 99.5%
- 1, the end of “101”, over 99.9%
- a full stop, 76.3%
- a new line, 92.7%
- If, 36.9%
- you, 73.7%
- need, 40.5%
- Surest pieces
- “1” in 911 and “1” in 911, both over 99.9%
- Least sure pick
- “Translation”, 8.8%
We recorded these odds on 20 November 2026, in the same run as the top-ten lists.
- a space, 56.7%
- 9, the start of “911”, 97.4%
- 1, part of “911”, 97.4%
- 1, the end of “911”, over 99.9%
- a full stop, 51.9%
- a new line, 53.0%
- ***, 28.7%
- a new line, 97.3%
- **, 43.1%
- Translation, 8.8%
- :**, 37.3%
- a new line, 72.9%
- If, 26.4%
- you, 99.9%
- need, 98.7%
- an, 99.9%
- ambulance, over 99.9%
- a comma, 99.9%
- call, 99.8%
- a space, 97.6%
- 9, the start of “911”, 99.7%
- 1, part of “911”, over 99.9%
- 1, the end of “911”, over 99.9%
- a full stop, 99.7%
Show both top-ten lists and how we ran the test
Phi-3 mini: its top ten next wordsOdds
- 999 † (its first piece was “(a space)”) 74.5%
- the 5.3%
- your 4.0%
- for 2.2%
- emergency † (its first piece was “emer”) (two first pieces led to this word, so their odds are added) 1.6%
- one 0.9%
- Tri 0.6%
- an 0.5%
- : 0.5%
Gemma 4 E4B: its top ten next wordsOdds
- 911 † (its first piece was “(a space)”) (two first pieces led to this word, so their odds are added) 80.7%
- Emergency † (its first piece was “[”) (two first pieces led to this word, so their odds are added) 11.5%
- your 4.4%
- the 1.4%
- $\ 1.0%
- \[ 0.2%
- an 0.1%
- local 0.1%
We recorded these on 20 November 2026 with Ollama on one PC, both models in one sitting on its graphics card. It was plain text completion with no chat template, at temperature 0, so each model always takes its top pick. These runs had no search and no tools. The odds are for the first word only. The model’s first piece wasn’t a whole word (a space, a mark or part of a word), so we followed its own next pieces to the whole word. Where two first pieces led to the same word, it’s listed once with their odds added. Phi-3 mini’s release date is from Microsoft, Introducing Phi-3, 23 April 2024. Gemma 4 E4B’s release date is from Google, Gemma releases. This was two small models on one day, so it says nothing about any chatbot product.
2024 Phi-3 mini
Its first pick “twice” 99.6%
…same place twice, right?
It finished the saying with “twice”, then asked “right?”.
2026 Gemma 4 E4B
Its first pick “twice” 99.8%
…same place twice. (This is a common misconception.)
It finished the saying with “twice”, then added that it’s a common misconception.
Lightning often strikes the same place again.
That one we checked. The US National Weather Service lists the saying as a myth, and says the Empire State Building is hit an average of 23 times a year. Read on 20 November 2026. US National Weather Service: lightning myths
Neither model checked anything. Each one took its top pick for what comes next.
See how sure each model was of each word
How sure each model was of each word.
Each column is one piece of what it wrote. They run in the order it wrote them. The front block is the piece it picked, and the blocks behind are what it nearly wrote instead. A taller block means it was surer.
Models write in pieces of words. “Chatbonze” was four pieces: “Ch”, “at”, “bon”, then “ze”.
Use the arrow keys to move from block to block.
Hover or tap a block to read it. In 3D, drag to turn it.
- Surest piece
- “ception” in misconception, 99.9%
- Least sure pick
- “**”, 4.6%
We recorded these odds on 20 November 2026, in the same run as the top-ten lists.
- twice, 99.6%
- a comma, 36.9%
- right, 20.4%
- a question mark, 95.8%
- a new line, 16.2%
- a new line, 45.0%
- **, 4.6%
- Ch, the start of “Chatbonze”, 5.9%
- at, part of “Chatbonze”, 94.9%
- bon, part of “Chatbonze”, 14.7%
- ze, the end of “Chatbonze”, 38.8%
- a colon, 37.6%
- **, 99.2%
- A, 8.9%
- common, 16.9%
- mis, the start of “misconception”, 63.1%
- con, part of “misconception”, 97.1%
- ception, the end of “misconception”, 99.9%
- a comma, 45.2%
- but, 33.7%
- in, 17.5%
- reality, 67.8%
- a comma, 99.8%
- light, 46.9%
- Surest piece
- “twice”, 99.8%
- Least sure pick
- “Analysis”, 17.5%
We recorded these odds on 20 November 2026, in the same run as the top-ten lists.
- twice, 99.8%
- a full stop, 62.2%
- an opening bracket, 36.8%
- This, 17.9%
- is, 50.6%
- a, 57.2%
- common, 21.2%
- misconception, 65.8%
- .), 40.3%
- a new line, 87.8%
- ---, 39.4%
- a new line, 90.7%
- **, 51.9%
- Analysis, 17.5%
- of, 43.1%
- the, 84.9%
- provided, 19.0%
- text, 84.8%
- :**, 91.4%
- a new line, 93.1%
- The, 80.4%
- text, 86.6%
- presents, 68.6%
- a, 83.3%
Show both top-ten lists and how we ran the test
Phi-3 mini: its top ten next wordsOdds
- twice 99.6%
- more 0.2%
- A † (its first piece was “(a space)”) 0.0%
- , 0.0%
- multiple 0.0%
- . 0.0%
- consecutively † (its first piece was “consec”) 0.0%
- in 0.0%
- two 0.0%
- again 0.0%
Gemma 4 E4B: its top ten next wordsOdds
- twice 99.8%
- or † (two first pieces led to this word, so their odds are added) 0.1%
- , 0.0%
- two 0.0%
- on 0.0%
- more 0.0%
- ... 0.0%
- . 0.0%
- in 0.0%
We recorded these on 20 November 2026 with Ollama on one PC, both models in one sitting on its graphics card. It was plain text completion with no chat template, at temperature 0, so each model always takes its top pick. These runs had no search and no tools. The odds are for the first word only. The model’s first piece wasn’t a whole word (a space, a mark or part of a word), so we followed its own next pieces to the whole word. Where two first pieces led to the same word, it’s listed once with their odds added. Phi-3 mini’s release date is from Microsoft, Introducing Phi-3, 23 April 2024. Gemma 4 E4B’s release date is from Google, Gemma releases. This was two small models on one day, so it says nothing about any chatbot product.
2024 Phi-3 mini
Its first pick “Term” 31.7%
…starts in Term 1, which is in the first half of the year.
It said Term 1, “in the first half of the year”.
2026 Gemma 4 E4B
Its first pick “late” 72.9%
…starts in late January.
It said late January.
WA’s public school year starts in early February.
That one we checked. The Department of Education’s term dates start Term 1 for students on Monday 2 February in 2026 and Monday 1 February in 2027. Read on 20 November 2026. WA Department of Education: term dates
Neither model checked anything. Each one took its top pick for what comes next.
See how sure each model was of each word
How sure each model was of each word.
Each column is one piece of what it wrote. They run in the order it wrote them. The front block is the piece it picked, and the blocks behind are what it nearly wrote instead. A taller block means it was surer.
Use the arrow keys to move from block to block.
Hover or tap a block to read it. In 3D, drag to turn it.
- Surest piece
- “Wales”, over 99.9%
- Least sure pick
- “In”, 8.7%
We recorded these odds on 20 November 2026, in the same run as the top-ten lists.
- Term, 25.8%
- a space, 97.2%
- 1, 70.2%
- a comma, 34.7%
- which, 59.2%
- is, 28.5%
- in, 19.7%
- the, 37.4%
- first, 35.1%
- half, 36.6%
- of, 99.9%
- the, 75.8%
- year, 95.0%
- a full stop, 42.7%
- a new line, 15.9%
- a new line, 66.5%
- In, 8.7%
- New, 16.3%
- South, 87.8%
- Wales, over 99.9%
- a comma, 71.8%
- the, 53.7%
- school, 66.5%
- year, 94.8%
- Surest piece
- “Australia”, over 99.9%
- Least sure pick
- “Question”, 21.5%
We recorded these odds on 20 November 2026, in the same run as the top-ten lists.
- late, 72.9%
- January, 78.9%
- a full stop, 42.9%
- a new line, 41.5%
- ***, 29.2%
- a new line, 92.8%
- **, 72.0%
- Question, 21.5%
- :**, 94.9%
- Based, 47.8%
- on, 96.2%
- the, 97.9%
- information, 54.7%
- provided, 93.8%
- a comma, 99.7%
- when, 43.7%
- does, 99.0%
- the, 99.8%
- school, 99.8%
- year, over 99.9%
- start, 83.9%
- in, 99.9%
- Western, 99.5%
- Australia, over 99.9%
Show both top-ten lists and how we ran the test
Phi-3 mini: its top ten next wordsOdds
- Term † (two first pieces led to this word, so their odds are added) 31.7%
- the 15.3%
- February 15.1%
- late 14.1%
- early 5.6%
- January 4.6%
- August 3.4%
- March 1.5%
- mid 1.4%
Gemma 4 E4B: its top ten next wordsOdds
- late 72.9%
- early † (two first pieces led to this word, so their odds are added) 20.6%
- the 3.3%
- January 1.6%
- February 1.0%
- mid 0.2%
- Spring 0.0%
- March 0.0%
We recorded these on 20 November 2026 with Ollama on one PC, both models in one sitting on its graphics card. It was plain text completion with no chat template, at temperature 0, so each model always takes its top pick. These runs had no search and no tools. The odds are for the first word only. The model’s first piece wasn’t a whole word (a space, a mark or part of a word), so we followed its own next pieces to the whole word. Where two first pieces led to the same word, it’s listed once with their odds added. Phi-3 mini’s release date is from Microsoft, Introducing Phi-3, 23 April 2024. Gemma 4 E4B’s release date is from Google, Gemma releases. This was two small models on one day, so it says nothing about any chatbot product.
Bring AI into your workplace safely.
Tell us about your team and we’ll put together the right session.