AI Risk Board

Explore

Pages

The AI Risk Klog

A risk-centric Knowledge discovery Log of articles, blog posts, podcasts, and observations in the journey to understand and adopt AI-related technology.

Filippo Marino

Keeping up with the breakneck rate of Artificial Intelligence (AI) innovation since the first public release of ChatGPT is practically impossible while holding a steady job. Everyone is struggling to develop a reasonably accurate and stable mental picture of AI, let alone of emerging applications, risks, and opportunities.

The bias of anything included herein reflects my

personal⁠

and

professional⁠

interest and expertise in the domain of

human risk judgment and decision-making (RJDM)⁠

and risk intelligence.

⁠

Highlights #6: The Inevitability of Multi-modal Battlefield AI

6/14/2023

The focus of this knowledge discovery log remains that of AI applications in the intelligence analysis and decision support functions, all from the perspective of a risk JDM and mitigation professional. (I need to remind myself of this here and there as the emergence of fascinating AI-related conversations and applications only grows by the day - quickly pulling us in unrelated directions as we experiment with methods and tools with measurable value-add.)

The past couple of weeks have convinced me that an LLM-driven battlespace decision-support system appears increasingly inevitable if we consider results like

Voyager’’s “embodied agent”⁠

that can move and act freely and purposefully in a simulated (Minecraft) or real environment. It’s an exceptional result that combines distinct tasks and tactical goals within a larger strategic objective. On the heels of

Meta’s Cicero project⁠

and the growing

multi-sensory/multimodal⁠

data ingestion and

training approaches⁠

, it seems inevitable that one such system will soon be substituting or, at minimum, complementing current military technology.

The most interesting articles and papers this week:

Highlights #5: Transformers, Agents, and Toxicity.

5/14/2023

The avalanche of AI-related news and innovations didn't slow down much these past two weeks. From a technical standpoint, the most consequential announcements we saw involve the increasing number and power of ChatGPT/LLMs

plugins⁠

and

transformers agents⁠

- along with

Anthropic’s release of a 100K context window⁠

- extending (rather, exploding) these models’ power and potential. Early discussions about

multi-sensory⁠

multimodal⁠

and multi-model AI (with the obvious AGI implications) are becoming increasingly real, fast.

Our

AI Risk Register⁠

holds as these remain within the classes and scenarios that have been articulated for some time. That being said, all our

Risk, Uncertainty, and Decision-making (RUD) conversations⁠

on this subject were (astonishingly) accurately and elegantly summed up in an

article by Jonathan Haidt and Eric Schmidt in The Atlantic⁠

Here’s how

Clarity⁠

would summarize the article:

brown and white smoke on brown rock formation

Highlights #4: AutoGPT, Jailbreaks, ...

4/18/2023

This past week’s release of

AutoGPT⁠

was followed - as expected and without delay - by the first (and profoundly deflating) deployment of a

humanity-destroying sequence⁠

, and accompanied by various

ChatGPT jailbreaks⁠

. None of this is surprising or beyond the scope of the

risks in our register⁠

and, in fact, it confirms the critical issues of innovation speed and anticipated lack of self-restraint associated with this technology.

Many of us started experimenting with either localized LLM deployment, with open-source projects like

OpenChatKit⁠

ColossalChat⁠

, and

Databrick’s Dolly⁠

, or by bridging ChatGPT and proprietary content with LangChan (

per our last post⁠

). But if all that is too much for you now, you can still get a sense of what a ‘chat’ with a specific/recent public or private document feels like with

ChatPDF⁠

, which does a terrific job.

If the concept of synthetic relationships fascinates you as well, you may want to hear this

No Priors podcast episode⁠

with Noam Shazeer of

Character⁠

. If you, like us, are approaching this subject from a risk intelligence perspective, you’ll find a certain naiveté about the entire venture (especially around min. 17), which is both disarming and terrifying at the same time.

Highlights #3: Prompts, Search vs Chat, 3rd Party Apps,...

4/10/2023

"It's the Prompt, Stupid!" And indeed it seems AI-Whisperers and

Prompt Engineers are the new Data Scientists⁠

. From overnight marketplaces (

PromptBase⁠

PromptHero⁠

ChatX⁠

, and so on) to Youtube classes, private training programs, and free, customizable, no-code apps like

Pickaxe⁠

(which I found to be a great way to get some practice with structured prompting.)

One more Washington Post story about this:

Tech’s hottest new job: AI whisperer. No coding required⁠

⁠

The two biggest commercial races (at least, for the the next few months) are search integration and proprietary content 'ingestion'. On the former (B2C-first; how to transition from search behavior to a conversation) Bing has taken a bit of a lead on Google (despite what is probably the least imaginative UX solution of the decade), but others like

Neeva⁠

and

Perplexity⁠

are trying to reinvent the search box, and they are worth checking out. On the latter front, the ChatGPT

'Copilot' integration with Microsoft 365⁠

and

Azure⁠

, along with

Salesforce's Einstein GPT⁠

are likely to be carry the widest and most immediate business benefits.

Highlights #2: Control, LLMs Limits, New/Small Models, ...

4/3/2023

Among the more frequent AI risk debates is that of control. On one extreme is the monopoly and dominance argument - one in which the number of state and commercial entities with the know-how and resources to develop and own such technology is increasingly small, and the race winner may achieve unprecedented hegemony, be it economic or geopolitical. On the other end, is the concern with unrestricted access to code, algorithms, and virtual machines capable of unimaginable harm. How the pendulum swings seems to change by the week, but if there is anyone with an opinion worth listening to is Matey Zaharia, CTO of

Databricks⁠

, who in

this episode of the No Priors podcast⁠

seems to suggest small models may be capable of achieving greater accuracy with smaller training sets than anticipated (hello VC, Are you listening?)

Directly related to the above, is the enduring debate about the strengths and weakness of Large Language Models (LLMs) vs Symbol Manipulation approaches, and I can't think of a more useful read than Gary Marcus' "Deep Learning Is Hitting a Wall" essay:

https://nautil.us/deep-learning-is-hitting-a-wall-238440/⁠

⁠

In this terrific Making Sense episode Sam Harris speaks with Stuart Russell and Gary Marcus about recent developments in artificial intelligence, the limitations of Deep Learning the long-term risks of producing Artificial General Intelligence (AGI), and of course, the control problem.

https://podcasts.apple.com/us/podcast/making-sense-with-sam-harris/id733163012?i=1000603166437⁠

⁠

Early Analytical Experiments

4/1/2023

Can ChatGPT facilitate structured (risk) intelligence analytical efforts? By its own response, it can. Yet, by now, we know that might be just an articulate and probabilistic answer. So we're putting some of these to the test, and getting very interesting results. Before stepping into specific methods, what is most impressive is the OpenAI’s capacity to perform content analysis on the fly:

Entity extraction (people, organizations, etc.)

Relationship mapping

⁠