Generative AI
- Generative AI
- A gentle intro to the explosion of LLMs with ChatGPT (2022) and others:
- ChatGPT is not "true AI.",
by AI scientist
Michael Wooldridge,
May 17, 2023.
- Written when ChatGPT launched, with the GPT-3 model.
- These models are massive:
"OpenAI didn't train GPT-3 on desktop computers.
They used very expensive supercomputers containing thousands of specialized AI processors, running for months on end."
- GPT-6 Astra (2026)
was trained at OpenAI's Stargate data center
in Texas.
It was trained on
over 100,000 NVIDIA H100 GPUs.
- Article discusses whether LLMs are "prompt completion"
(too harsh, even then)
or at the edge of true AI intelligence
(probably too excited).
- My position is that generative AI has made me consider how much of subconscious
human thought may be like LLMs.
More than we have thought.
OpenAI
OpenAI API
Click to run World:
Chat with GPT 3.5 at
Ancient Brain.
Chat with GPT 3.5 model.
This is an open source JavaScript World to call the
OpenAI API to
talk to
GPT 3.5 (2022).
GPT 3.5 was the model used by
ChatGPT
when it started in 2022, launching the AI revolution.
- In these earlier GPT models there was a parameter called "Temperature".
-
This is "temperature" in the softmax output.
Click to run World:
Chat with GPT 5.6 at
Ancient Brain.
Chat with GPT 5.6 model.
This version talks to
GPT 5.6 (2026).
This is the model ChatGPT uses as at 2026.
- Generative AI is superb at generating code.
- This has led to huge debate as to whether there will be no jobs for software engineers now,
or whether jobs in software will simply evolve.
- Generative AI is especially good at short, self-contained functions. (Small context window.)
Things Stack Overflow
was brilliant for.
- It is more of a challenge to build large systems with many files and inter-connecting parts,
and have AI keep track of all the context.
-
There is also a fundamental issue of risk
if you deploy production code that has never passed through a human brain.
Not just "code reviewed".
We mean, the human brain thought about it, invented it, removed parts, added parts, changed things,
and while doing so considered hacking injections, divide by zeros, null strings, and so on.
-
Vibe coding
is the idea of getting an AI to write your code by talking to it in human language.
True "vibe" coding is where you do not really look at or check the code, but just keep speaking to the AI
to fix any issues.
- There is much debate on what "vibe coding" will do to the software industry:
- On the one hand, it allows "creatives" who cannot code, but are good on user interfaces,
aesthetics, etc., to mock up beautiful prototypes.
- On the other hand,
vibe coding by people who cannot code
is likely to generate code with many subtle bugs, issues, security holes, etc.
- 2025 survey
found that 45% of AI generated code samples
introduced
OWASP Top 10 security vulnerabilities
into the code.
-
You, who can code, may be spending your career cleaning up after vibe coders!
