Skip to content
Skill Trees: Gamifying The Hard Things by Steph Piper
- A list of skills
- Each area has a series of skills that can be colored in.
- Design
- Hexagons are good
- Can be done in any order, hard to connect meaningfully
- Simple, flexable milestones
- Reception
- First on was 3d printing & modeling
- Tested on makerspace student staff members. Good to identify gaps
- Benefits
- Reduce imposter syndrome or on the other size overconfidence
- Target areas for improvement
- Online on git – https://github.com/sjpiper145/MakerSkillTree
- How to make a skill tree
- Flexibility, not too cost restrictive, globally applicable
- Peer reviewed
- Final skill tree and translation
- Book – The Learning Game by Ana Lorena Fabrega
- Beta testing book of a collection of these skills.
- Good published through “Make: Magazine”
- 68 tiles per tree, 1020 skill tiles in the book
- Tips for writing
- Continue to evolve and improve
- Do own illustrations was huge time saver from the publisher
- Confidence in your work. The publisher will only do the final publishing
- Looking to fill the gaps
- Working on a kids version of the book
The Token Wars: Why not everything should be open by Kathy Reid
- The Token Wars
- A resource conflict fought through technical, social and legal means
- What is a token?
- An atomic unit of text taken from a larger collection called a corpus
- text -> subwords tokens -> vectorization
- Transformer architecture
- Word embeddings capture semantic closeness of words
- Scaling up to billions of tokens
- Train the relationships between tokens based on all the text
- The value of tokens and token economics and the actors in the token wars
- Are the a public good?
- No the are rivalrous either excludable or non-excludable
- LLMs in 2024 were trained on 4 orders of magnitude data than 5 years ago.
- Estimated 60-160 trillion tokens on the public web and some LLMs are trained on close to all of those
- Synthetic Data especially low quality slop is polluting the Internet
- Scrapers pick this up and train on it, concern about Model Collapse ( like a photocopy of a photocopy). Reduces the diversity of what it will produce.
- Key actors in the token wars
- Individual content creators
- Included in corpus without permission
- Platforms with user-generated content
- Seeking to get paid for their content ( eg Reddit deal with OpenAI )
- Archival Institutions
- Australian National Film and Sound Archive: Maintain Trust, Transparancy, Create Public Value
- Private Companies
- Anthropic: Model Context Protocol
- The AI Companies
- Have used fair-use. Although some countries don’t have those
- Companies blocking the common crawl
- Governments
- Having trouble balancing interests
- Token Tactics – Protecting your token treasure
- Data poisoning
- Blocking bots and scrapers
- Data Sovereignty
- Futures
- Hunt for more tokens
- Better ways to block/prevent
- Better understanding of the alateral damage of the resource conflicts