diff --git a/_posts/2023-12-27-generative-ai-predictions.md b/_posts/2023-12-27-generative-ai-predictions.md new file mode 100644 index 0000000..2b586ae --- /dev/null +++ b/_posts/2023-12-27-generative-ai-predictions.md @@ -0,0 +1,98 @@ +--- +title: "Generative AI Predictions" +author: Greg +layout: post +permalink: /2023/12/generative-ai-predictions/ +date: 2023-12-27 09:41:25 -0500 +comments: True +licence: Creative Commons +categories: + - technology +tags: + - generative-ai + - llms + - gpt + - chatgpt + - llama + - news + - predictions +--- + +The potential of Generative AI goes far beyond Large Language Models (LLMs like ChatGPT) being able to answer trivia questions, write generic essays, and code using established patterns. The true value of Generative AI is its ability to organise information and then reason. + +## Extracting semantic information and reasoning + +Generative AI models can extract semantic meaning from documents. Document types include; + +- text (articles, chat conversations, etc) +- images (photos, x-rays, scanned documents, etc) +- audio (voicemails, soundscapes, heart beats, etc) +- video (TV shows, live footage, dash cams, etc) +- numerical (server logs, medical results, etc) +- timespace (coordinates, timestamps, etc) +- multimodal (a mix of model types; a video with coordinates and a timestamp, an image with a description, etc) + +Consider the task for collecting an individual's tax deductible receipts. Receipts could be text documents in the form of an SMS message, image documents from photographed receipts, or multimodal documents from PDFs attached to emails. A generative AI model with access to the individual's SMS messages, photo library, and email inbox could read each document, extract the semantic information, and reason if each document is a tax deductible receipt. This use case will save humans millions of toilsome hours each year. + +## Predictions + +### Teaching the news + +Journalists access primary sources (press releases, interviews with individuals, etc) and generate news stories for an audience. The audience may be the general public in the case of nightly news or more narrow in the case of articles in industry publications. The journalist must make assumptions on what the audience knows in order to efficiently communicate the news. + +Consider this [Al Jazeera article about China’s influence in the Solomon Islands](https://www.aljazeera.com/economy/2023/12/5/in-the-solomon-islands-a-us-agencys-struggles-hint-at-chinas-influence). The journalist wrote this story for a global audience and reasonably assumed the average reader might not know that the small nation of the Solomon Islands is in the Pacific Ocean. While useful for the average reader, this additional context is inefficient for readers familiar with the Solomon Islands. + +I'm a part-time acrobatics teacher with Diana, my partner. Diana is a teacher by trade and an important lesson I have learned from her is efficiently communicating the information for the student to succeed. In the case of partner acrobatics, it means knowing the level of the student, understanding their learning style, observing where the student can improve, and delivering tailored cues to help the student succeed. + +Generative AI can deliver this same level of attention for news delivery. Consider a platform that consumes local town hall meetings, press releases, interviews, social media posts, weather, etc. and critically evaluates the information to effectively communicate the pertinent information for you. For instance: + +- Your local councillors are voting on putting a new park 2km from your house. It's unlikely to pass, here is your local councillor's email address +- The car you drive is having a voluntary recall for the seat warmers +- The UN has released new climate change predictions which align with peer reviewed articles from climate experts. However your federal representative continues to refute the findings on social media and received $50K from pro fossil fuel lobby groups in the last 12 months. + +In this reality, the journalist role would change from writing articles to conducting interviews, investigative journalism, and designing these news platforms. + +### Consciousness via a neural network of concepts + +It's unlikely that a generative AI model will gain consciousness any time soon. However, a system using generative AI that extracts concepts from documents, links the concepts together with weighted connection, and uses this knowledge network to reason may achieve a level of consciousness. I will expand on this concept in an upcoming post. + +### Generative AI auditing + +Generative AI's ability to extract semantic meaning from documents will be useful for identifying previously unnoticed patterns in communications. AI auditing will surface corruption, exonerate individuals, and rewrite history. + +## Local generative AI + +Locally run generative AI models like [Mistral's Mixtral](https://mistral.ai/news/mixtral-of-experts/) are preforming well at reasoning tasks. Mixtral running on my gaming PC correctly answers the *if it takes 3 hours for 3 towels to dry, how long does it take for 9 towels to dry* question which demonstrates high reasoning skills. + +```json +{ + "model": "mixtral:latest", + "created_at": "2023-12-28T02:39:18.966611317Z", + "messages": [ + { + "role": "user", + "content": "If it takes 3 hours for 3 towels to dry outside, how long does it take for 9 towels to dry outside?" + }, + { + "role": "assistant", + "content": " Given that drying conditions are similar, we can apply the same time duration for drying 9 towels as it took for 3 towels. This is because the drying process happens concurrently for all towels rather than sequentially. Therefore, if 3 towels take 3 hours to dry, then 9 towels would also take 3 hours to dry under the same conditions." + }, + "done": true, + "total_duration": 9265104510, + "load_duration": 975707, + "prompt_eval_count": 39, + "prompt_eval_duration": 1940182000, + "eval_count": 83, + "eval_duration": 7322496000 +} +``` + +This completely local process ran in 7 seconds and cost less than $0.0001 in electricity. Subscription cloud generative AI platforms like OpenAI's ChatGPT may succeed however advertising funded platforms like Microsoft's Bing will struggle as interactions with Bing basically become conversations with a salesperson. + +## Frameworks + +Using existing open source models it is possible to develop all the frameworks outlined in this blog post. However, open source models will continue to improve with better reasoning abilities, larger context windows, and increased efficiency. + +## Dangers of generative AI + +It's important to highlight the dangers of generative AI. 2024 is an election year in the United States and we are likely to see further weaponization of generative AI. Tools already exist to maintain trust online. For instance, public-key cryptography where a private key can authenticate a message was written by a particular authour. In fact, this blog post is signed using public-key cryptography. In order to impersonate me, an attacker would need to access my securely held private key[¹](/2022/03/trust/). Unfortunately these tools are not widely adopted or understood. No politician or organization should be making public statements without using public-key cryptography irregardless of the threats of generative AI. Generative AI is just an accelerator of the existing threats. diff --git a/_posts/2023-12-27-generative-ai-predictions.md.asc b/_posts/2023-12-27-generative-ai-predictions.md.asc new file mode 100644 index 0000000..5ee3c6d --- /dev/null +++ b/_posts/2023-12-27-generative-ai-predictions.md.asc @@ -0,0 +1,117 @@ +-----BEGIN PGP SIGNED MESSAGE----- +Hash: SHA256 + +- --- +title: "Generative AI Predictions" +author: Greg +layout: post +permalink: /2023/12/generative-ai-predictions/ +date: 2023-12-27 09:41:25 -0500 +comments: True +licence: Creative Commons +categories: + - technology +tags: + - generative-ai + - llms + - gpt + - chatgpt + - llama + - news + - predictions +- --- + +The potential of Generative AI goes far beyond Large Language Models (LLMs like ChatGPT) being able to answer trivia questions, write generic essays, and code using established patterns. The true value of Generative AI is its ability to organise information and then reason. + +## Extracting semantic information and reasoning + +Generative AI models can extract semantic meaning from documents. Document types include; + +- - text (articles, chat conversations, etc) +- - images (photos, x-rays, scanned documents, etc) +- - audio (voicemails, soundscapes, heart beats, etc) +- - video (TV shows, live footage, dash cams, etc) +- - numerical (server logs, medical results, etc) +- - timespace (coordinates, timestamps, etc) +- - multimodal (a mix of model types; a video with coordinates and a timestamp, an image with a description, etc) + +Consider the task for collecting an individual's tax deductible receipts. Receipts could be text documents in the form of an SMS message, image documents from photographed receipts, or multimodal documents from PDFs attached to emails. A generative AI model with access to the individual's SMS messages, photo library, and email inbox could read each document, extract the semantic information, and reason if each document is a tax deductible receipt. This use case will save humans millions of toilsome hours each year. + +## Predictions + +### Teaching the news + +Journalists access primary sources (press releases, interviews with individuals, etc) and generate news stories for an audience. The audience may be the general public in the case of nightly news or more narrow in the case of articles in industry publications. The journalist must make assumptions on what the audience knows in order to efficiently communicate the news. + +Consider this [Al Jazeera article about China’s influence in the Solomon Islands](https://www.aljazeera.com/economy/2023/12/5/in-the-solomon-islands-a-us-agencys-struggles-hint-at-chinas-influence). The journalist wrote this story for a global audience and reasonably assumed the average reader might not know that the small nation of the Solomon Islands is in the Pacific Ocean. While useful for the average reader, this additional context is inefficient for readers familiar with the Solomon Islands. + +I'm a part-time acrobatics teacher with Diana, my partner. Diana is a teacher by trade and an important lesson I have learned from her is efficiently communicating the information for the student to succeed. In the case of partner acrobatics, it means knowing the level of the student, understanding their learning style, observing where the student can improve, and delivering tailored cues to help the student succeed. + +Generative AI can deliver this same level of attention for news delivery. Consider a platform that consumes local town hall meetings, press releases, interviews, social media posts, weather, etc. and critically evaluates the information to effectively communicate the pertinent information for you. For instance: + +- - Your local councillors are voting on putting a new park 2km from your house. It's unlikely to pass, here is your local councillor's email address +- - The car you drive is having a voluntary recall for the seat warmers +- - The UN has released new climate change predictions which align with peer reviewed articles from climate experts. However your federal representative continues to refute the findings on social media and received $50K from pro fossil fuel lobby groups in the last 12 months. + +In this reality, the journalist role would change from writing articles to conducting interviews, investigative journalism, and designing these news platforms. + +### Consciousness via a neural network of concepts + +It's unlikely that a generative AI model will gain consciousness any time soon. However, a system using generative AI that extracts concepts from documents, links the concepts together with weighted connection, and uses this knowledge network to reason may achieve a level of consciousness. I will expand on this concept in an upcoming post. + +### Generative AI auditing + +Generative AI's ability to extract semantic meaning from documents will be useful for identifying previously unnoticed patterns in communications. AI auditing will surface corruption, exonerate individuals, and rewrite history. + +## Local generative AI + +Locally run generative AI models like [Mistral's Mixtral](https://mistral.ai/news/mixtral-of-experts/) are preforming well at reasoning tasks. Mixtral running on my gaming PC correctly answers the *if it takes 3 hours for 3 towels to dry, how long does it take for 9 towels to dry* question which demonstrates high reasoning skills. + +```json +{ + "model": "mixtral:latest", + "created_at": "2023-12-28T02:39:18.966611317Z", + "messages": [ + { + "role": "user", + "content": "If it takes 3 hours for 3 towels to dry outside, how long does it take for 9 towels to dry outside?" + }, + { + "role": "assistant", + "content": " Given that drying conditions are similar, we can apply the same time duration for drying 9 towels as it took for 3 towels. This is because the drying process happens concurrently for all towels rather than sequentially. Therefore, if 3 towels take 3 hours to dry, then 9 towels would also take 3 hours to dry under the same conditions." + }, + "done": true, + "total_duration": 9265104510, + "load_duration": 975707, + "prompt_eval_count": 39, + "prompt_eval_duration": 1940182000, + "eval_count": 83, + "eval_duration": 7322496000 +} +``` + +This completely local process ran in 7 seconds and cost less than $0.0001 in electricity. Subscription cloud generative AI platforms like OpenAI's ChatGPT may succeed however advertising funded platforms like Microsoft's Bing will struggle as interactions with Bing basically become conversations with a salesperson. + +## Frameworks + +Using existing open source models it is possible to develop all the frameworks outlined in this blog post. However, open source models will continue to improve with better reasoning abilities, larger context windows, and increased efficiency. + +## Dangers of generative AI + +It's important to highlight the dangers of generative AI. 2024 is an election year in the United States and we are likely to see further weaponization of generative AI. Tools already exist to maintain trust online. For instance, public-key cryptography where a private key can authenticate a message was written by a particular authour. In fact, this blog post is signed using public-key cryptography. In order to impersonate me, an attacker would need to access my securely held private key[¹](/2022/03/trust/). Unfortunately these tools are not widely adopted or understood. No politician or organization should be making public statements without using public-key cryptography irregardless of the threats of generative AI. Generative AI is just an accelerator of the existing threats. +-----BEGIN PGP SIGNATURE----- + +iQIzBAEBCAAdFiEESYClA57JitMYg1JBb8nUVLEJtZ8FAmWM9dMACgkQb8nUVLEJ +tZ9xmBAAg07aPTf9ZWznQ3hgdxDsUokhcLREAqJFBP5o5W/2tvVps+qTnrSGBSUZ +LeAYP9JrJdKCp+N5sa+Ljg9GD+xp6uh+MVchDvG3eu5DEBBYWN3t9+hesJauqxup +oPLT6VHTxui2GpAUlOxVF+N82nolf4NM9R+xvPRpNK71gdIeKke9cNSQdZe9Odb6 +2rFCsnuQtkwDkBTQ6XW3cFI+ncoyu2xkqj0qfN+clXo5ckulApYRUZtCKEroW9ar +1Wn6kPPEsIfGIyVYTEB500JvqsXGRioC9ZdE2p+tHlI36HDvCs0NVaTW/KLoM4s6 +PQSyjkLKKIeT98zxYbjUwbG5+3q4ssbVUNuFkkW+wpgyrjxbkAt4tbCEkTdoLrEE +yPJpr8c1/h4uX9gIYV5tN5eakgbZVj2d4JI++RiIBxxZhRxRtlalt/L1jUPIB9Fg +M3qe5FdUQCXTXuq2iHfedG9MstRbuM4+j3v8VVlRggF8+V/a7olACGPBGfq1OiUn +9Ai9qojldXmxDgFIafoTto0WpTs/GjH62WXsHqf2FOZ9fnFYFEuYB1IK4wyUAgry +MZkw83VZ2ME+fjJXlynNS6sIsfiB8JfalqgbhV3OQC4v8BbK/3pyi+X8cFw9lMnm +KybH/bOSVgUP3v5o9CmGZmSEAwZ2ARGVl5C4cltKGHzkLtpYeS8= +=tUKx +-----END PGP SIGNATURE----- diff --git a/codex.md b/codex.md index 8c9164d..8ebe904 100644 --- a/codex.md +++ b/codex.md @@ -278,6 +278,10 @@ Subconscious reactions that I have noticed; - Humans will develop new common language to discuss recent populism (bubbled advocates, post-truth politics, etc) - Car fitness will become a trend as people utilize their self driving commute +### Before 2026 + +- Hyper-personalized news based on generative AI will exist + ### Before 2030 - Human driven vehicles will be uncommon (driven partly by the cost of insurance and an increase in the value of human life) @@ -287,6 +291,7 @@ Subconscious reactions that I have noticed; - ~~AIs will easily pass Turing tests~~ - 🟢 Success: In 2023 large language models like ChatGPT 4 can pass the turing test in all but a few edge cases - AI audits will rewrite history by discovering patterns in open & closed data +- AI audits will uncover many instances of political and corporate corruption - Biometric tracking will allow for early detection of communicable diseases greatly reducing their transmission potential & prevalence - Critical thinking skills will improve as legacy media fades (alternatively the world will be destroyed by trolls) - Critical thinking skills will improve with [regulation of algorithms](https://en.wikipedia.org/wiki/Regulation_of_algorithms) diff --git a/codex.md.asc b/codex.md.asc index 7174350..138c0d9 100644 --- a/codex.md.asc +++ b/codex.md.asc @@ -281,18 +281,26 @@ Subconscious reactions that I have noticed; - - Humans will develop new common language to discuss recent populism (bubbled advocates, post-truth politics, etc) - - Car fitness will become a trend as people utilize their self driving commute +### Before 2026 + +- - Hyper-personalized news based on generative AI will exist + ### Before 2030 - - Human driven vehicles will be uncommon (driven partly by the cost of insurance and an increase in the value of human life) - - Traffic will be dramatically reduced by the small number of vehicles using traffic aware navigation (Uber drivers, self driving cars, delivery trucks, etc) - - Deurbanisation will occur as proximity to cities becomes less important with new technology like self driving vehicles and virtual reality - - Pedestrian will reclaim downtown areas from cars -- - AIs will easily pass Turing tests +- - ~~AIs will easily pass Turing tests~~ + - 🟢 Success: In 2023 large language models like ChatGPT 4 can pass the turing test in all but a few edge cases - - AI audits will rewrite history by discovering patterns in open & closed data +- - AI audits will uncover many instances of political and corporate corruption - - Biometric tracking will allow for early detection of communicable diseases greatly reducing their transmission potential & prevalence - - Critical thinking skills will improve as legacy media fades (alternatively the world will be destroyed by trolls) - - Critical thinking skills will improve with [regulation of algorithms](https://en.wikipedia.org/wiki/Regulation_of_algorithms) - - At least one of the Detroit Big Three (Ford, GM, & Chrysler) will fail to adapt and be acquired or go bankrupt +- - The majority of information accessible and accessed on the Internet will be generated by large language models (LLMs) like ChatGPT +- - LLMs will amplify mistakes & misinformation in unintended feedback loops when new generations of LLMs are accidentally trained on the output of previous generations ### Before 2040 @@ -323,17 +331,17 @@ Subconscious reactions that I have noticed; - - An extraterrestrial society will deem earthlings worthy and reveal themselves to us -----BEGIN PGP SIGNATURE----- -iQIzBAEBCAAdFiEESYClA57JitMYg1JBb8nUVLEJtZ8FAmQIxZYACgkQb8nUVLEJ -tZ99QA/9GssqFJPhnsf94bD3ZY3xOpCRaMvkQ7piJiY8Ayd+0ghvYMrKajQpulZg -jB5i3H9zDkPoUYn3tVpwR4d+esUOvM3uzPxVI7VYM2VNi52tGilw3lvbR0EPHZC8 -0uwUeRLaSSWRszjIYyOTvO7yWroAwABa7NfcXeg8MHdO/BE/3C9fpVt5UIggoedP -3O/JePZvhTUrrEDhr3g2kQwCZC15DyR4qYHm8Y9mkGX+j8Jh6NBeJ/NARfZhImGR -qWvikKkg6Z5FcByZHiI67+YUvPVpSEvYshMY37lu96yzyUuYUrokvDP1W2M5Lkti -+c9xaHVVp+Jm8wnTaaPKFRsRM0Hokm2Y8UYM/WaShgimte7W+end+nuQGDXwJG7x -cP1irrSRRav5/SLQnan5EPnKMyo6B5tWADqse5TclxiQsUWKofLjKsiM/Q+ilC8t -HaGF2CU35FigQqXLnZzjlCC0JHAwM7rTK7wc18v4KRt2TrA9SvGmymAEFA0fr5Km -TggZdaXvHvggEJz35DKZIHQ1MXOrHdKkjA+JER/Fo1k0uD4dWHq9SJRCgK2AXljr -8uxoh+ugz8qqMVRRoYClbAgpLBv+fK+1rgtBdrM80CzeDTM8vhH0OjlyBCW+pPnG -aYE6zUpzMl8/y8+tCDYRsl3GCHX7CTleX0fPVOZSlQKFtSEcZrM= -=u1zF +iQIzBAEBCAAdFiEESYClA57JitMYg1JBb8nUVLEJtZ8FAmWM9AoACgkQb8nUVLEJ +tZ/PTg/7Bz+pi7Wh8/TKbUs03nqv5TFu27dq2jIZIzP1EPxozBzChtypg5Uuqjo9 +liykm8IFwkNe0XKbXs56EN5XWHPtfRmnEmsjIKHiWNGmXAt7eQWZwr99s4uKBV9H +iDGBAMLjUdZ2etkeJthIXjzTJRJPk9C+bgZcJI7faxHPdDW3O7nK3vbyTq+6dFBP +5hvH3K2YiEqZHxJqu7qZb4U9o8YUTZpYFTR43KlvYZ0mUg5YWmKhyHfVVEo6aJ8P +VVtlipG92SYTbF93vv2Ru+bZHsTtdklTZ8SOBA8I+NLlbhfOCWBhnc+oC7502PYd +Pb//+KqBZOStyQvwDucOsOAwgELSDm1cW2yTtps1yRkuq7MHebDShZPFLi54JQs5 +jymZ2Jc3sgNAlYS0uKsuz91kVUMx8xM/FXcYFViM4BJjBj7Ejajn/udvLpmmxyFG +oilOReTA6XNZFkLsctcu/qlS44m6U8ORX58SOuZh4Yz1EyLiJ9n43uC1k1R3aZfw +6ZBii7/lXFtKFvLk30ocD0NRscVbMJCT+OK/tqrYIfdiDtkuC0WiD8YJkUnvzQKf +6F/PdgXdw8X57xh2stuiTQwTmEWlC8EV2cxCjy7QNxas6ivCXs4QzeF2xRhFIKv5 +ajwA4dsaNSnKeUussOYQ9uyuEv3FXSC5FzXfPy5CPWettytswT8= +=0ez3 -----END PGP SIGNATURE-----