Clippy the Maximizer

A talk about the lesser known dangers of chat GPT and other generative models.

Threats

Chaff: flooding society
- Bullshit asymetry principle. AI Chaff principle is 2-3 orders of magnitude bigger.
- Slows everything down, a penalty on everyone doing anything.
Identity impersonation
- Especially friends and relatives
- This is in the wild already
- Dangers of creating synthetic profiles with or informed/inspired by real data
- Chaffing someone’s identity by creating similar mock profiles to make their real profile hard to find and trust
- Click Monkeys: It can exist now at even larger scale
Collecting information about you from relatives and contacts
- Combining this info via enrichment
Psyops and manipulation
- Sentiment analysis
- Goal directed agents
- Synthesizing people, voices, content, entire identities and websites for false histories and legitimacy
- Replika is a good example of this threat, think of this happening at scale
Propaganda and psyops
- Create profiles by combining data from multiple real entities
- Leaving fake data trails from synthetic individuals
- Could I disguize this as advertising and get VC funding to finance this
- Most advertising is about manipulating populations, so yes
- Pysops as a service: basically Facebook but much more dangerous
Unreliable generation
- How many pieces of sound are in a cloud?
- Broken is dangerous, but occasionally working is worse
- Accuracy of self driving cars article (6 nines of reliability)
- How accurate is ChatGPT, is it even 2 or 3 nines?
Changing motivatios of companies
- Companies change the operation of services frequently, AB testing as well
- Companies shut things down constantly
- Shaky foundation on which to build other things
- AI Supply chain issues: data sources, training personnel, etc.
- What if the company gets sold to an adversarial party?

Pleading. Read this. chatGPT has no concept of what lying is. It has no concept of fact, because the data has not been properly validated amongst other things. What is happenning is highly sophisticated probability testing.

It is burbling. Follow the links. OpenAI themselves describe it as a work in progress. Two or three iterations down it may be more reliable With a bit of luck…. See my comments below for the original post and other supporting detail. Tread carefully.

Source: post by Alan Woods

Can we trust anything from ChatGPT?

This carefully researched article demonstrates just how dangerous generative Large Language Models are.

They dream up the sequence of tokens following tokens in a way that produces highly plausable streams of words, fabricating plausible sources with complete disregard for any reality.

This is because the LLM has no interest in what documents provided any information. Indeed it is not created or intended to be able to do this.

Its design has no capability to link its token patterns that it is learning to any one of the vast number of training documents.

Those who have created this technology do not understand linguistics, the science of languages. They believe that intelligence can be obtained by this dreaming technology just by using ever larger models, but, as we are seing, this i not true and can never be the case.

This is the greatest danger and risk from all the hype surrounding current neural net based AI.

Source: post by Richard Self

Amongst other concerns it seems chatGPT has democratised the writing of malicious code. Its not really a specialist skill anymore as it will generate such code on request. Tears before bed time. — Alan Woods