CoBot Systems

26/09/2023

On games, winning, and cheating.

Would you think that Wordle players cheat? Why would they? And yet, "Many Wordle users cheat to win, says mathematics expert" 😲

It seems there's a five-letter word describing what many players of the wildly popular Wordle puzzle do daily as they struggle to find a target word within six tries.

20/09/2023

Can AI catch lies?

We can tell when we're being lied to about 47% of the time ... we might as well just flip a coin. So we generally believe that what we're told. Our defense mechanism is that we avoid listening to liars.

Financial analysts listen to CEOs to help them analyze the CEO's company. These analyses guide investors and move the stock market. Hyde’s team has built an AI that determines CEOs' lies with 84% accuracy.

Let the evolutionary race begin! We can imagine the predator/prey evolution of analysts+TruthAI versus CEOs+AvatarAI 😀

Ref:

Research Summary Organizations are punished by analysts and investors when material deceit by their CEO is uncovered. However, few studies examine analysts' responses to deceptive CEOs before their ...

19/07/2023

We hear about high prices for new vehicles in the USA. Cox compares the vehicle-prices to income.

According to the Cox Automotive/Moody’s Analytics Vehicle Affordability Index. The number of median weeks of income needed to purchase the average new vehicle in June 2023 was steady at 43.0 weeks.

The Cox Automotive/Moody's Analytics Vehicle Affordability Index shows new-vehicle affordability was unchanged in June.

14/07/2023

VanMoof e-bike customers are at the risk of their bikes becoming bricked if the company shuts down.

VanMoof is a Dutch e-bike success story. It was founded in 2009 by two brothers to make the "perfect city bike": custom built & premium. The bicycles look slick, employ sound effects, have a built-in battery, etc.

VanMoof bikes depend on VanMoof’s servers to operate. The bikes can be controlled by a smartphone. This connection is encrypted with a key obtained from VanMoof’s servers when logging onto the mobile app. If VanMoof’s servers go offline, customers cannot get this key, meaning they cannot communicate with their bike.

The electric bike company raised close to $200M, but now closed all its service points. For bike customers, they could face the risk of their bikes being bricked if they don't take prompt action.

06/07/2023

Version 13.3 of Wolfram Language has a whole subsystem around LLMs. Wolfram|Alpha lets anyone use natural language to get questions answered. Now with LLMs it’s possible to use natural language to start defining potential elaborate computations.

The most immediately visible LLM tech in Version 13.3 is Chat Notebooks. Go to File > New > Chat-Enabled Notebook and you’ll get a Chat Notebook that supports “chat cells” that let you “talk to” an LLM.

Wolfram's Version 13.3 includes a subsystem around LLMs plus new functionality in integrals, special functions, finite fields, computational geometry, live highlighting in visualizations, AR & VR deployment. See examples for those and dozens of other enhancements.

30/06/2023

The S-shaped curve of the Prospect Theory subjective value function naturally emerges as the optimal solution to the problem of value perception.

The good reasons behind reference-dependence

15/06/2023

McKinsey identified 63 generative AI use cases spanning 16 business functions that could deliver total value in the range of $2.6 trillion to $4.4 trillion in economic benefits annually when applied across industries

Generative AI’s impact on productivity could add trillions of dollars in value to the global economy—and the era is just beginning.

15/06/2023

The illusion of moral decline
Adam M. Mastroianni & Daniel T. Gilbert, Nature (2023) https://www.nature.com/articles/s41586-023-06137-x

Morality is not actually declining. Studied across a 55-year span from 1965 to 2020, people’s reports of the current morality of their contemporaries were stable over time. On average, the year in which the survey was conducted explained less than 0.3% of the variance in responses, and in almost all cases it explained less than 1%.

If morality has not declined, then why do people think it has?

First, numerous studies have shown that human beings are especially likely to seek and attend to negative information about others, and mass media indulge this tendency with a disproportionate focus on people behaving badly. As such, people may encounter more negative information than positive information about the morality of ‘people in general’, and this ‘biased exposure effect’ may help explain why people believe that current morality is relatively low.

Second, numerous studies have shown that when people recall positive and negative events from the past, the negative events are more likely to be forgotten, more likely to be misremembered as their opposite, and more likely to have lost their emotional impact. This ‘biased memory effect’ may help explain why people believe that past morality was relatively high.

Working together, these two phenomena can produce an illusion of moral decline. Specifically, biased exposure to information about current morality may make the present seem like a moral wasteland, biased memory for information about past morality may make the past seem like a moral wonderland and when people in a wasteland remember being in a wonderland, they may naturally conclude that the landscape has changed.

We show that the perception of moral decline is pervasive, perdurable, unfounded and easily produced, and suggest that this illusion has implications for research on the misallocation of scarce resources, the underuse of social support and social influence.

14/06/2023

Marc Andreessen on AI: If you don’t agree with the prevailing niche morality that is being imposed on both social media and AI via ever-intensifying speech codes, you should also realize that the fight over what AI is allowed to say/generate will be even more important – by a lot – than the fight over social media censorship.

AI is highly likely to be the control layer for everything in the world. How it is allowed to operate is going to matter perhaps more than anything else has ever mattered. You should be aware of how a small and isolated coterie of partisan social engineers are trying to determine that right now, under cover of the age-old claim that they are protecting you.

In short, don’t let the thought police suppress AI.

There's a full-blown moral panic about AI right now. But the real risk is losing the race to global AI technological superiority.

10/06/2023

The White House released the national AI R&D strategic plan and is soliciting public input on critical AI issues. You can weigh in by June 7.FACT SHEET: Biden-H

10/06/2023

The strategic intent to make AI so cheap and easy that everyone can use it runs up against the desire to add safeguards. "If only wealthy hospitals can take advantage of AI systems, the benefits of these technologies will not be equitably distributed." vs five core protections i the US AI policy:
1. Safe and Effective Systems: You should be protected from unsafe or ineffective systems.
2. Algorithmic Discrimination Protections: You should not face discrimination by algorithms and systems should be used and designed in an equitable way.
3. Data Privacy: You should be protected from abusive data practices via built-in protections and you should have agency over how data about you is used.
4. Notice and Explanation: You should know that an automated system is being used and understand how and why it contributes to outcomes that impact you.
5. Alternative Options: You should be able to opt out, where appropriate, and have access to a person who can quickly consider and remedy problems you encounter

Safeguards seem to add costs to AI, which goes against making AI cheap and widespread. If we can convert the problem of safeguards into a problem of effectiveness, we can get to the virtuous cycle where AI becomes cheap and effective while honoring the safeguards.

Each AI owner is already internally incentivized to build an AI-checker to ensure that the AI is effective (i.e., it does what it's supposed to do). That addresses the requirement for effective AI, where unsafe AI is handled as being ineffective. Let's call this sort of AI "Effective AI".

Algorithmic discrimination and data privacy work against the effectiveness imperative of the AI owner. Denial of data would typically be used by the AI as part of its algorithm (what does it mean for my decision that Tom has requested that xyz datapoint should not be used in his case?). Whether by AI or not, discrimination is the basis of decision-making. Locating and combating illegal discrimination can and should rest with the agencies responsible for eliminating it, because those social-good agencies can convert an otherwise open-ended problem to a well-defined set of algorithms. This approach, again, aligns with the natural need for using an effective AI, in this case to further the interests of the social-good agencies. Let's call this AI to be "Social Alignment AI" that its owners will want to be efficient and effective.

Notice and Explanation stems from being fair. As currently worded, it can create an expensive arms race as malicious players use explanations to understand and win against the AIs. A black box that's doubly guarded by Effectiveness and Social Alignment can be sufficient to meet the need for fairness.

Alternative Options deals with the ethics of not trapping people. A real "way out" should enable review and redress, stemming from being just. These are difficult matters anyway, not just for AI, and need to be addressed as a set of "Justice AI" that seems to be an aspect of Alignment AI. The problem is that an unjust AI decision can be trivially easy to find-and-solve or devilishly hard, possibly rife with false-positives and false-negatives. This class of "Social Alignment AI" is likely to use triage approaches.

Separating the concepts of Effectiveness and Social Alignment AI will, I think, provide both cost-effective and safeguarded AI. Each AI owner is incented to make its AI efficient and effective. This ecosystem requires a market-making player to enable each Effectiveness AI to check its outputs against the relevant set of Alignment AIs so that it can constantly locate and eliminate misalignments.

The White House released the national AI R&D strategic plan and is soliciting public input on critical AI issues. You can weigh in by June 7.FACT SHEET: Biden-H

22/05/2023

Can decision-making benefit from an AI driven co-pilot?

The Vice President and CEO of Innovation at Microsoft, Jason Wild, predicted that every job will be transformed by an existence of an AI co-pilot, driven by large language models (LLMs) such as ChatGPT (see https://www.laprensalatina.com/microsoft-every-job-will-have-an-artificial-intelligence-copilot/).

At CoBot Systems, we have been building OR-DA algorithms for decision support for years. For us, these are "co-pilots". They are used for decision-making by car dealers in the USA (see https://www.frogdata.com).

So decision-making does benefit from an AI driven co-pilot because it provides decision-makers with the embedded intelligence of an analytics practitioner who has learned by immersion in the business context, and that gets updated as we (the system providers) find ways to improve it.

Using LLM for these business decision agents is non-trivial. Think of it this way:
1. It's a known category of problem, and an optimal solution exists. This would use classical OR-DA.
2. It's a new category or an unsolved problem. When used in a business context, it is stated as developing situation awareness (akin to battlefield awareness) and then path-finding in explore/exploit cycles. Again, classical methods provide firm guides for both awareness (including anomaly detection) and experiment-assessment cycles.

If we include LLMs in the decision-support algorithm, the analytics practitioner is still required to build it. That practitioner can use an LLM as a co-pilot. Analytics practitioners acting as Decision Coaches (h/t Dr. Steve Barrager and Dr. John Milne) would also have a role to play in helping business decision-makers use the decision co-pilots because of the range of concerns to be handled. Those Decision Coaches could use an LLM as a co-pilot. So there are three kinds of LLM-co-pilots here:
1. Decision Co-Pilot = classical OR-DA co-pilots used by business decision-makers, augmented by LLM only where appropriate.
2. Decision Co-Pilot Maker = LLM to help the Decision Co-Pilot builder (an OR-DA practitioner). Such a Maker can democratize the role of analytics practitioners.
3. Decision-Coach Co-Pilot = LLM to guide the Decision Coach (an OR-DA practitioner) who helps business decision-makers use the Decision Co-Pilot.

Are we getting to the point of having an INFORMS BoK LLM that could serve these three uses?

This is in a discussion thread from INFORMS.org (see https://connect.informs.org/discussion/can-decision-making-benefit-from-an-ai-driven-co-pilot).

Friends,The Vice President and CEO of Innovation at Microsoft, Jason Wild, predicted that every job will be transformed by an existence of an AI co-pilot, drive