AI 2024 update: Large Action Models, AlphaGeometry, Graphcast, Mixtral, Prompt Injection

Oliver got Chris Booth and Charles Phiri back along with Peter Gostev who has recently moved to a Head of AI role at Moonpig to discuss recent developments in the AI space.

They discuss recent developments in AI, with a focus on new models like AlphaGeometry, GraphCast, and Mixtral. The participants also discuss benchmarking of models, prompting frameworks like LangChain and LangGraph, security issues like prompt injection attacks, integrating AI capabilities into enterprise architectures, prompt engineering. data architecture etc! They talk about the need to treat AI systems more like humans in terms of governance, testing, and managing unpredictability. A wide-ranging conversation touching on many aspects of AI progress and challenges.

Peter also fills us in on his various AI exploits and research into the security and safeguards of popular GenAI models as mentioned on his LinkedIn: https://www.linkedin.com/in/peter-gostev/

00:00 Welcome
00:20 Introducing Peter Gostev
01:06 Episode overview
01:54 Latest AI Models
02:30 Large Action Model (LAM)
03:00 Mistral Mixtral
03:36 Mistral Medium Model
05:40 AlphaGeometry
07:33 Graphcast
08:25 Energy use of HPC vs Predictive Model
11:08 Meta - Llama 3 and massive GPU investment
13:07 Demand for GPUs will only increase
14:10 More to come from Deepmind?
14:45 AI arms race and FilteringAI
15:12 GPU numbers used for training
16:43 Nvidia getting out of Gaming?
17:08 Neuromorphic computing
17:28 Memristors
20:15 Hardware scaling challenges
21:00 Prompting Frameworks
21:51 Langchain and Langgraph
23:52 Knowledge and Reasoning
26:17 Where do we draw the line on understanding?
29:00 Throwback to UNIONS in SQL
29:50 Data Engineering
30:10 Vector vs Graph technologies
31:20 Combination of the 2
32:30 InferGPT and Neo4J Cypher
33:01 Self building approaches
34:03 Layers and Iterations
35:20 Learn what works through controlled failure
35:59 Benchmarking
38:30 How do you test these things?
39:28 Drift by design
40:30 This gets into philosophy
40:50 Treating these things more like humans
43:40 These things can be socially engineered
44:00 Peters work on prompt exploitation
46:05 More complex models = more attack vectors?
47:39 Breaking the 2 layers of filtering
50:30 Does this need regulation?
52:40 How much do companies need to worry about this?
54:00 Multiple language prompts
55:30 AI is a component not the silver bullet
57:20 Data Science and Software Engineering merging
59:07 When would we trust this technology?
1:00:20 Opening up access to knowledge
1:01:28 Just training to pass tests
1:03:01 Conclusions - what should be prioritised
1:08:00 Short term vs long term thinking

Show notes will be updated soon including...
- URLs for speakers
- Links to Models
https://mistral.ai/news/mixtral-of-experts/
- Langchain and Langgraph
- InferGPT etc

Видео AI 2024 update: Large Action Models, AlphaGeometry, Graphcast, Mixtral, Prompt Injection канала Architect Tomorrow