Difference between revisions of "User:Darwin2049/ChatGPT4/Conclusions Synthesis"
Darwin2049 (talk | contribs) |
Darwin2049 (talk | contribs) |
||
Line 96: | Line 96: | ||
These include: Tool-X, Tool-Y. One then uses these tools to prepare the communications channel from one's web site to one's account with OpenAI. | These include: Tool-X, Tool-Y. One then uses these tools to prepare the communications channel from one's web site to one's account with OpenAI. | ||
<BR> | <BR> | ||
'''''<Span Style="COLOR:WHITE; BACKGROUND:BLUE">Fine-Tuning.</SPAN>''''' | '''''<Span Style="COLOR:WHITE; BACKGROUND:BLUE">Fine-Tuning.</SPAN>'''''<BR> | ||
Where that technology enabled them to leverage their existing competitive position then they were able to gain advantage by improving their ability to occupy their particular market position or improve their ability to deliver results as a result of improved operational efficiency. <BR> | Where that technology enabled them to leverage their existing competitive position then they were able to gain advantage by improving their ability to occupy their particular market position or improve their ability to deliver results as a result of improved operational efficiency. <BR> | ||
We should expect to see the same process play out with CP type capabilities. In order to make this less theoretical and more concrete a review of the existing tools that are now available will help put this perspective in context. When we review landscape of tools several come immediately to the fore. These include several LLM's as well as several GAN based systems. The LLM's | We should expect to see the same process play out with CP type capabilities. In order to make this less theoretical and more concrete a review of the existing tools that are now available will help put this perspective in context. When we review landscape of tools several come immediately to the fore. These include several LLM's as well as several GAN based systems. The LLM's are led currently (March 2024) by Google and Microsoft and FaceBook. <BR> | ||
'''''Google''''' is the umbrella organization under which '''''[https://www.wikiwand.com/en/Google_DeepMind DeepMind]''''' operates.<BR> | |||
'''''Microsoft''''' is the umbrella organization under which '''''[https://www.wikiwand.com/en/OpenAI OpenAI]''''' operates.<BR> | |||
'''''Meta (FaceBook)''''' is the umbrella organization under which '''''[https://www.wikiwand.com/en/OpenAI LLama]''''' operates.<BR> | |||
In the case of Google we find DeepMind at the forefront of artificial intelligence tools. In the case Microsoft we find OpenAI establishing a powerful presence in the AI marketplace with its (formerly) open source ChatGPT4 system. In the case of Meta (FaceBook) we find their Llama 2 system. Another contender to be considered is Anthropomorphic. In each case these offerings that these enterprises offer enable a knowledgeable user to access and interact with remarkably powerful AI systems.<BR> | |||
In the case of GAN based systems we find Dall-E and MidJourney. These powerful tools enable a user to provide a text prompt that is then used to generate a visual representation based upon the terms and parameters that they user provides. Recently (Feb. 2024) OpenAI has announced its Sora system which offers a user the ability to enter a text prompt and receive a stunningly lifelike three dimensional moving video. Currently there is a time limit for the duration of the video which is about sixty seconds. Going forward this will doubtless increase. Yet even with Sora's capabilities as they are currently understood we see an AI system that is capable of generating entirely lifelike output that is indistinguishable to the average observer from actual reality. | In the case of GAN based systems we find Dall-E and MidJourney. These powerful tools enable a user to provide a text prompt that is then used to generate a visual representation based upon the terms and parameters that they user provides. Recently (Feb. 2024) OpenAI has announced its Sora system which offers a user the ability to enter a text prompt and receive a stunningly lifelike three dimensional moving video. Currently there is a time limit for the duration of the video which is about sixty seconds. Going forward this will doubtless increase. Yet even with Sora's capabilities as they are currently understood we see an AI system that is capable of generating entirely lifelike output that is indistinguishable to the average observer from actual reality. | ||
<BR> | <BR> |
Revision as of 03:46, 2 March 2024
Overview
Interfacing/Access - Synthesis, Conclusions.
- Adeptness with CG4. Skills with this new capability will be comparable to taming fire.
Specifically, risks that either involve or actually are:
CG4, its derivatives and brethren will introduce consequences just as transformative; we therefore note that:
- CG4 CP's are Dual Use. Just like fire, CG4 will be used for both constructive as well as destructive purposes;
- Pre-literate vs. Post-literate Societies. it is a powerful enabler for leveraging a collection of people toward forming a self sustaining, much more advanced social structures and organizational units;
- New Horizons. CG4 will enable the recognition, articulation solutions to problems that less adept societies could not;
- On Bringing a Knife to a Gun Fight - Prospects. pre-CG4 social polities will be at a severe disadvantage comparable to the encounter between conquistadors vs indigenous New World natives. At minimum we should expect to see CG4 class CP's being used as:
- Self Contained Tools. An increasing recognition that CG4-Class tools can serve as capabilities in and of themselves; that this is evident can be seen when a query is posed to CG4 to translate a tract of text to another language, or provide a synopsis of it; these are all intrinsic capabilities of the tool... in and of itself;
- "Bridging" Mechanisms. These will serve as CP connectors to other more specialized capabilities such as IBM Watson or CYC Corp's CYC common sense knowledge base system; it is important to keep clearly in mind that humans operate at ultra-low bandwidth when conversing; considerable information is embedded in context rather than explicitly stated; mechanism such as CG4 and its peers are capable of interacting at electronic speed; therefore can exchange enormous volumes of information almost instantly;
- Composite Cognitive Prosthetics. Multiple instances configured to be autonomous agents based upon CG4 can be collected together to form "societies of mind"; initial versions might have many to tens of instances; later variations might range into the hundreds or thousands of agents; theoretically the only practical upper limit to how many agents can be connected together is limited by the physical hardware available at the time;
- Extended Composite CP Systems. Marketplaces offering ad hoc or persistent composite capabilities will proliferate. These will offer features and capabilities that can only be guessed at.
How Different Groups Can or Will Interact with this New Capability?
Different groups will interact with this new capability using the primary interfacing mechanism known as prompting. Further interactions will be performed using fine-tuning.
Prompting.
Direct Access.
Free. Currently CG4 can be accessed at not cost. However limitations apply. Using the free access modality one can develop one's skill at formulating prompts. At the current time, prompts and prompting are the primary means of accessing and interacting with CG4. However like any new software tool with considerable power and capability the more proficient the user is at formulating prompts the higher the quality of the results that will be available.
Subscription.
Anyone interested in exploring the capabilities of CG4 can sign up with an account. By virtue of having an account a user will have access to a broader range of possibilities and performance that are not available to those using the free access pathway. An extended subscription pathway is available wherein a second user can access an account as a team member.
Local Access.
There will be cases wherein a user might want to explore how to capitalize on CG4's capabilities but may have data that may be categorized as private. As such the possibility that it might become publicly accessible is ruled out. The way to work around this problem is to use a local version of CG4 that runs on one's own equipment. There are several tools that can be used to accomplish this objective. These include GPG4ALL, PrivateGPT and H2O.
The GPT4ALL is the simplest and most direct way of acquiring familiarity with an LLM and developing proficiency in controlling it. GPT4ALL can be downloaded at not cost to one's own Windows or Apple environment, installed and activated. This is open source software and therefore expect that there will be small inconsistencies in the setup process.
PrivateGPT is a comparable open source LLM. It is a somewhat more sophisticated tool. It also requires considerable technical skill. It can be brought up on one's own environment but doing so requires being versed in the world of UNIX commands.
H20 is another relatively sophisticated tool. It also requires considerable technical competence in order to bring it up to operational status.
GPT4ALL.
PrivateGPT.
H2O.
Remote Access.
User Website.
Should one prefer to enable access to CG4 via one's own website there are tools available that can enable this objective. The basic elements that are needed are a) one's own website, b) an account with CG4, c) two user interface tools that facilitate and mediate communications between the user's website and one's account on OpenAI.
These include: Tool-X, Tool-Y. One then uses these tools to prepare the communications channel from one's web site to one's account with OpenAI.
Fine-Tuning.
Where that technology enabled them to leverage their existing competitive position then they were able to gain advantage by improving their ability to occupy their particular market position or improve their ability to deliver results as a result of improved operational efficiency.
We should expect to see the same process play out with CP type capabilities. In order to make this less theoretical and more concrete a review of the existing tools that are now available will help put this perspective in context. When we review landscape of tools several come immediately to the fore. These include several LLM's as well as several GAN based systems. The LLM's are led currently (March 2024) by Google and Microsoft and FaceBook.
Google is the umbrella organization under which DeepMind operates.
Microsoft is the umbrella organization under which OpenAI operates.
Meta (FaceBook) is the umbrella organization under which LLama operates.
In the case of Google we find DeepMind at the forefront of artificial intelligence tools. In the case Microsoft we find OpenAI establishing a powerful presence in the AI marketplace with its (formerly) open source ChatGPT4 system. In the case of Meta (FaceBook) we find their Llama 2 system. Another contender to be considered is Anthropomorphic. In each case these offerings that these enterprises offer enable a knowledgeable user to access and interact with remarkably powerful AI systems.
In the case of GAN based systems we find Dall-E and MidJourney. These powerful tools enable a user to provide a text prompt that is then used to generate a visual representation based upon the terms and parameters that they user provides. Recently (Feb. 2024) OpenAI has announced its Sora system which offers a user the ability to enter a text prompt and receive a stunningly lifelike three dimensional moving video. Currently there is a time limit for the duration of the video which is about sixty seconds. Going forward this will doubtless increase. Yet even with Sora's capabilities as they are currently understood we see an AI system that is capable of generating entirely lifelike output that is indistinguishable to the average observer from actual reality.
In order to get some sense of the power of these tools can be facilitated if the user develops some familiarity with how to interact with them. Currently there are several ways in which to accomplish this. An avenue of choice is to use a ChatGPT-type capability that can be operated on one's own local machine. There are several tools that enable this. These include GPT4All, PrivateGPT and H2O. The knowledgeable user can also facilitate access to their personal website and have ChatGPT4 access the website. A user is then able to interact with ChatGPT from the perspective of an individual web site.
If we take the path of least resistance then the GPT4All is the choice of a pathway forward. It can operate on both a Windows or an Apple machine. It is significantly easier to install and get operating. The PrivateGPT requires significantly more technical skill in bringing online. It can be downloaded both to Windows or Apple environments. But does require technical competence to get working. The H20 is also not straightforward in terms of activating.
By way of example we include a partial printout of a dialog between GPT4All and a user. Using this approach a prospective researcher or user can gain insight into how prompting works and how to improve performance via fine-tuning.
- Innovators vs Laggards. It is not uncommon to encounter a friend, family member or colleague who is extremely adept at using the latest features of a smart phone or pad type computer; while on the other hand there are others who are baffled by the variety of features available on a common Android or iPhone and limit their usage to managing voice or text interactions; beyond that they basically waste the other 90%+ of the device's capabilities;
- Selective Usage. There are also those who find that their "smart" device can do more than send or receive communications but can perform many other capabilities; as a typical user might explore and employ those that are of greatest utility to themselves; but this simply reconfirms that these are actually general purpose computers; they simply have no moving parts, run on batteries and collapse the user interface onto a small color screen;
- New Day Same Routine. We should expect to see the same pattern repeat; those who are motivated to explore the new capabilities will excel at leveraging its capabilities; those who do not will remain perpetually playing catch up;
How Will or Can They Respond to its Existence and Presence. Recent studies have made plain that a powerful new technology will be adapted in record time if it is demonstrably a powerful leveraging tool; evidence of this can be seen in how quickly it is adapted.
- The CG4 "Uptake Curve". CG4 reached over one million users in a matter of days.
- Within Less Than Two Months. CG4 had reached 100 million users worldwide. This rapid uptake pace suggests that momentum has built that will drive further uptake by those less inclined to vigilance about the rapid pace of change that this technology brings;
- New Capabilities. An entire library of CG4 material has sprung into existence on how to use create new applications and extensions;
How will they be affected by this new capability.
- Personalized Tutors. CG4 provides a user with the ability to pose remarkably sophisticated questions on a broad range of topics; where the acquisition of a new skill or area of expertise is envisioned it is capable of creating a study guide with a complete syllabus that provides a pathway from even the most superficial grasp of a topic to its most advanced areas;
- Interactive Audio Books. Audio books will become passé; instead it will become preferable to be able to have a conversation with the book itself; were one to study the works of Konrad Nietzsche one would instead be able to effectively converse or otherwise "talk" with Nietzsche and discuss the various topics that the actual philosopher entertained when writing his works; we should expect to see variations on this capability wherein several legacy experts simulacra were instantiated and communications between them or existent users was enabled; a case in point might be enabling a conversation between Frank Herbert (author of Dune), Alan Watts and for instance, Jordan Peterson; a user initiating a session such as this would experience the two instances interacting and discussing with each other points of clarification; clearly even larger sets of knowledge areas can be brought to bear as well;
- Cross-Referenced Expertise Multiple insights will become available within a single interactive session; for instance were one to study the topic of Stoicism as described by Epictetus and his peers as described, one might be able to re-represent his insights into a compact symbolic logic representation; such a representation structure might be used to provide insight on family or personal matters by individuals trained as sociologists or psychotherapists; for a given psychotherapeutic setting, several points of view might be offered on a specific psychological issue; one perspective might derive from a Jungian approach to the problem; another perspective might derive from a Berne approach (transactional analysis); these two perspectives might "discuss and debate" with each other their respective positions and subsequently offer their recommendations on how to proceed;
- Individualized Advisors (professional and life coaches). Individuals who aspire to the top ranks incorporations have come to recognize that navigating a successful path to the top requires a range of skills including leadership, communication and narrative story telling. It is uncommon to find the full set of these skills in any specific individual. Enabling access to the insights of experts in each of these areas will become more accessible and available.
- Enhanced Collaborative Problem Solving. Existing WebEx, Skype or Zoom teleconferencing have become common in enabling voice and video online and real time discussion; stepping into a higher dimension we should expect to see individuals come to the fore each with their individualized CP tool set and offer access to their peer experts;
- Emergence of marketplaces for novel objects, products and services. CG4 class CP's are already spawning new products and services; one case in point can be seen with the availability of organizations offering vector database services to host extended memory for LLM's; the existing CG4 offers a broad range and depth of capability; for specific purposes however fine tuning is indicated; looking forward a short way down the road we should expect to see organizations offering their services or tools that facilitate the fine tuning task; the ability to fine tune a LLM will mean that there will be an explosion in the range, sophistication and depth of capability of LLM based tools;
- Novel and Inexplicable Disruptions Of service providers such as wireless or other online capabilities. The recent ransom-ware attacked that happened in Las Vegas in the US underscore the reality of malicious use of netcentric tools as weapons; as netcentric attacks were beginning to ramp up several years ago a favorite vector of attack was to use the distributed denial of service (DDOS) approach; in this approach hundreds or thousands of PC's were redirected to request a response to a specific web server; the requests were repeated endlessly until the hosting server was overwhelmed and was unable to respond; the result was that the website timed out; it became inaccessible; a variation on this approach can be used with comparable tools but targeted in a coordinated fashion at mobile phones; the result can be a combination traffic management attack combined with mobile phone attacks; that this is plausible is corroborated by the fact that users of such services as Tik Tok are able to geolocate a specific individuals location as well as launch other forms of disruption to that individual's mobile phone; were many individuals phones to be used as an attack vector then many people could find themselves trapped in a traffic gridlock and unable to reach others; the Israeli Pegasus tool has made very clear that it is fully capable of seizing control of any mobile phone on the planet;
- High Risk/Payoff Early Adapters. The early adaptors will explore the new terrain and gain the advantages that those who have mapped out the advantageous terrain is relative to the pitfalls. We should see a proliferation of tools that facilitate fine tuning, simplification of incorporation of vector databases that dramatically expand on a CP's ability to access a specific domain or set of knowledge domains.
Might the various access modalities that are available to one group have positive or negative implications for other groups.
- Early vs. Late Adapters. This pattern appears to repeat itself; therefore we should expect to see the same thing happen; those who can leverage these capabilities will define the new high ground; the rest will play catch up; or simply be left behind "in the past";
- Increasingly Chaotic Anecdotal Realities. Further fragmentation appears to be emerging as the new normal; deep fake news reports, conspiracy theories are likely to abound; new services will emerge comparable to Ground News wherein reporting sources will be ranked in terms of bias and reliability.
- Phenotypical Bifurcation. Currently the perceptible difference between individuals who are adept with a new technology vs. those who are not is typically not obvious; going forward the differences may remain somewhat obscure but the distance between the two categories are likely to widen to the point that one might as well identify the one group - early adapters and innovators as being members of a different phenotype compared to those who are slow to adapt; with the former group becoming adept at navigating the new reality in effect leaving the latter group to navigate with poor dexterity;
- Skills Explosion - Generalists & Specialists. New skill collection will emerge with the new environment; the explosion of the World Wide Web led to new categories of individual who were adept at web site creation, with the vast expansion of web sites the need for search emerged which led to organizations that organized and indexed the population of existing and new web sites; these refined their reason for existence by ranking and thereby promoting various sites; this led to the entirely new skill of search engine optimization (SEO); myriad new services have emerged such as virtual private networking, ad blocking and automatic notifications; the new environment will experience a comparable emergence of specialist skills that will facilitate usage of this new "fire";
Interface - Question. The question as posed is to answer is how will differing groups or communities make use of this new technology. As in, how might one group intrinsically gain new advantage relative to others. In the case of zero sum or non-zero sum situations how might this technology confer advantage (or disadvantage) by one group relative to others.
Interface - Answer. Large Language Models represent an expansion of the possible. Entirely new sets of possibilities are emerging as a result. Answering the question of how different groups will use this new technology and how they will gain or lose advantage will depend upon how they navigate their way into this new realm. As of early 2024 there were several pathways forward. These included:
- Prompt Engineering.
- Super-Prompting.
- Fine Tuning.
- Targeted Tools and Uses.
- Autonomous Agent Tools.
Cognitive Prosthetics. In order to provide context it might be helpful to consider that the world has seen massive advances in technology, science, industry and commerce as a result of the hardware and software tools that we have come to take for granted. The include such commonly available tools as Microsoft Office, image editing, video construction and many other highly specialized software tools. These tools enable the design and creation of incredibly sophisticated products such as massive ocean cruse liners as well as amazingly sophisticated nuclear powered submarines. In all cases these tools typically have specialized access mechanisms. They typically require hours to days of specialized training to become conversant in their use. Then weeks or months of hands on practice to develop expertise and competence.
Advanced Cognitive Prosthetics (APCs - First Generation). This latest generation of tools should be considered as representing advanced cognitive prosthetics. This is because the means of access allows for direct verbal interaction and communication. At face value this might seem like a huge leap. However it must be kept in mind that in order to capitalize on systems of this order of sophistication the level of sophistication of the user is crucial. Their ability to fully leverage the power of these systems is grounded on the fact that they possess deep understanding of what the system is capable of. The converse side of the equation is the reality that verbal communication is inherently low bandwidth. Despite that being the case the availability of a conversational ACP means that new possibilities for exploring the search for solutions is now at another level.
Advanced Cognitive Prosthetics (APCs Second Generation). Large Language Models (LLM’s) such as ChatGPT4, LLAMA or DeepMind are software tools that function as powerful prosthetics. They bear resemble to the most sophisticated 3D printer or CRISPR-CAS13 genomic editing tool. They impart unheard of versatility and power to a competent user. Therefore we view LLM’s and their peers are advanced cognitive prosthetics (ACP’s). We expect to see them being used to further enhance already existing processes or tasks in ways that have heretofore not been imagined.
Advanced Conformal Cognitive Prosthetics (ACCP). The current generation of LLM’s tools that enhance existing cognitive functioning and task completion in ways that are very comparable to having a knowledgeable assistant in performing an already known or defined task.
In this view a ACP's traits could be viewed from the perspective of a surgeon's glove. In this case it would be like enhancing the glove with a variety of unobtrusively incorporated supporting tools. Continuing with this metaphor one might imagine that the fabric of the glove has a built in scalpel, light, laser, magnifying lens, clamps and suturing tools.
Keeping the reality of different groups gaining or losing advantage we expect that existent interactions between varying communities will continue as before. The sophistication and pace of exchange will vary however.
New Day Same Problem. Different groups will seize this new technology and search for ways in which to further their existing positions. This means that existing operational or transactional approaches will be preserved but modified as this new technology’s new capabilities and limitations become increasingly evident.
Prompt Engineering. Repeated use of LLM has placed focus on the crucial role of being able to pose useful queries to the system. LLM's possess a great deal of knowledge that is generally opaque to a user. The specific forms of response or knowledge is intrinsically present in the system however eliciting relevant responses that illuminate that knowledge is challenging. Therefore we are seeing a rise in the category of users or super-users who construct highly focused prompts for specific or broad use.
Domain Query Experts. Prompt Engineering has already revealed an iteration when it comes to accessing and using a LLM. In the case of broad availability of open source LLM's several groups have emerged who place focus on formulating highly effective prompts. These prompts are described as being "Super Prompts". Typically these Super Prompts incorporate multiple refinement structures that direct the LLM to perform a query in highly sophisticated ways. Early prompts used what have been called chain of thought approaches.
A tree of thought is a more powerful form of prompting. It enables more sophisticated queries to be posed to an LLM. With the tree of thought a prompt would direct the LLM to perform its functioning in a manner that resulted in parallel actions being executed against the LLM.
forest of thought are even more powerful querying techniques that facilitate deeper access into the LLM's capabilities.
LangChain. This is a development framework that a very knowledgeable software developer would use to construct domain or task specific tools that use an LLM as a core element.
Recent developments have emerged that place focus on providing very sharp and targeted queries to a LLM. This new approach is characterized by what are known as Super Prompts. In this case a prompt is constructed
Darwinian Escalation Ladder. This will prove to be a continuing journey. This new technology is self replenishing and self extensible. Therefore answering the question of how different groups will interact as the modalities of interaction will constantly change.
Current Snapshot. Therefore at most answering this question can only be tentative and iterative. Any current characterization of how individuals or groups will use this new technology can only be momentary. Subsequent characterizations will require modification from insights that will be refined and developed over time.
Known. Therefore in order to move forward suggests approaching this question from a relatively obtuse starting point. By this we mean that the means of using or sharpening and focusing on the ability of this capability over time.
By way of example. These fall into a small group that suggest how it is currently being used. What can immediately be identified are the means whereby a LLM is initially trained. Subsequent iterations of use will reveal that generic LLM’s can have their performance improved. There are several means wherein further performance improvements are suggested by several venues. These arise as a result of how LLM’s function and are brought forward for broad or general use.
Categories. In subsequent sections of this document the reader will find a short list of known areas or activities which are known to be impacted by advances in the world of CP’s.
- Teaching/Training/Advising/Entertaining.
- Informing/Persuading.
- Analysis.
- Summarizing.
- Illuminating/Discovering.
- Subverting/Suborning/Disabling This is a live map of world wide net-centric threats. Looking forward we would expect to see the frequency and sophistication of these attacks to increase pace and tempo.
- Malicious Dual use ACP tools will be used for hacking or other forms of net-centric attacks.
Examples. Cases already exist where an existing trade, skill or profession have large populations of individuals who use their skills in these areas.
Local/Sequestered GPT Instances.
Deep Fake Tools.
- Voice. The ability to clone a person's voice with a high degree of fidelity and accuracy increases the range of constructive as well as malicious uses that this new technology can be applied to.
- Imagery. Tools such as MidJourney and Dall-E are capable of creating imagery that is indistinguishable from photographic imagery. The trend toward increasingly lifelike imagery will simply continue.
- Video. The ability to construct a video from any narrative is now possible with existing deep fake video editing software. Using tools that enable face swapping means that any person whose face has or can be captured in a public venue can merged with t he image of a prominent person such as a politician or cinematic figure. Moreover anything that the fake video providers might want that public person to appear to be saying can now be done seamlessly to the point that it is almost impossible to determine if that public figure actually said what was presented.
Fine Tuning.
- LoRa.
- Bulk User Curated.