symbolicAGI

You are not likely to have 17.03800786369594 crashes if your Tesla is driven on autopilot for 130 miles.

The M suffix on 7.63 indicates million.

130 miles / 7630000 = 0.000017038 expected accidents while driving on Autopilot to your Mum's. You would have to make 58692.334781078 such trips before an expected accident. Suppose each trip takes 3 hours. That is 20.100114651 years of continuous driving. Probably you would not have an expected accident visiting your mum even if driving that trip was your sole activity.

ChatGPT-4 is wonderful for composing regular expressions. Saves so much time when transforming various Java strings that arise in my work.

Today's big time savings came from this prompt: "Write a Java method that uses the Eclipse AST parser to create a simple markdown file showing the commented method signatures of a given Java class text file."

"electronic dance music about a solo inventor pursuing his life long dream of creating artificial general intelligence over many decades."

Making a playlist of the same prompt with oh so many music genres.

Consider LLM code generation. It would an ideal domain for simulated data.

For example, a version of popular software is released with new functions. A framework could simulate numerous plausible code snippets that exercise the new release, including edge cases. The code snippets would be amply commented, and proven to work in a test harness.

This method is the opposite of depending upon the internet for training data. And this method is being used to train AI robots, that have no available internet training data.

The point is that symbolic computation as performed by Cycorp was held back by the need to train the Knowledge Base by hand in a supervised manner. NNs and LLMs in particular became ascendant when unsupervised training was employed at scale.

Perhaps LLMs can automate in large part the manual operations of building a future symbolic knowledge base organized by a universal upper ontology. Considering the amazing emergent features of sufficiently-large LLMs, what could emerge from a sufficiently large, reflective symbolic knowledge base?

Doug Lenat, RIP. I worked at Cycorp in Austin from 2000-2006. Taken from us way too soon, Doug none the less had the opportunity to help our country advance military and intelligence community computer science research.

One day, the rapid advancement of AI via LLMs will slow down and attention will again return to logical reasoning and knowledge representation as championed by the Cyc Project, Cycorp, its cyclists and Dr. Doug Lenat.

Why? If NN inference were so fast, we would compile C programs with it instead of using deductive logical inference that is executed efficiently by the compiler.

Easiest way to guess a person's age is by the relative size of the ears, that never stop growing.

Tiny ears - teenager. Huge ears - senior.

https://pubmed.ncbi.nlm.nih.gov/18653105/

X (formerly Twitter) has two sorts of feeds that are chosen.

1. "Following", which is chronological.

2. "For you", which is algorithmic.

For those whose motivation is tracking news events, the chronological view is most important. Yesterday's news may not matter for example.

Otherwise the algorithmic feed is awesome. In my case I want Ukraine news and only the Western side presented. The algorithmic feed gives me interesting posts from accounts that I do not necessarily follow, and omits the Russian point of view - which is perfect.

Indeed. The common answer for scaling compute is the recruit more hardware.

Tesla has a vast amount of video data to feed into its vision --> action model. Training on Nvidia hardware takes too long. Dojo fixes this problem by a novel SIMD architecture.

Being on the spectrum entails downsides as well as super powers.

Elon is a jerk, Elon is a genius, Elon is the best custodian for the vast capital he controls.

https://abilitycentral.org/article/autism-quick-tips-top-8-t...

Benchmarks I've run with a Ryzen 7950x, 128 GB RAM with Nvidia GeForce 3060 12 GB VRAM show a slowdown less than half when not using the GPU, with LLama.cpp as the inference platform and various ggml open source models in the 7B-13B parameter range.

The Ryzen does best with 16 threads, not the 32 it is capable of, which is expected due to it having 16 CPU cores.

The Stackoverflow research initiatives per generative AI are very good examples of what some enterprises can do to improve user-input text context on websites and applications.

1. Adapt user text to a standard format. 2. Correct spelling and grammar mistakes. 3. Detect inconsistencies during the input process.

Wonderful innovation to use handy buttons that invoke specific prompts, reducing user's cognitive load from an otherwise empty chat input box.

The graph reminds me of concept maps used in GOFAI knowledge representation and acquisition.

Obvious dimensions for improvement include sets of button-prompts specialized for various LLMs including open source ones. Plus the set of buttons could be dynamically configured per user per topic per graph using reinforcement learning or a recommendation engine technique.

Sign me up.

Exponential growth of recursively self-improving AI happens simultaneously in several dimensions.

1. Better algorithms, more efficient, faster, more capable. 2. Better hardware, for example bespoke chips for each particular common computation. Could figure out molecular computing as Drexler predicted. 3. Recruit more computer capacity. Every networked computer could be used by a well-funded AI willing to pay enough. 4. Applying AI to exponentially increasing quantity of tasks by progressive decomposition of tasks into subtasks. Each worked in parallel as needed. 5. AGI would interact with the physical world with robotics that it designs, manufactures and operates.

If this is correct, then we could expect what AI-safety folks name "FOOM". Link: https://www.lesswrong.com/posts/t8dtuYXPhwt93PstB/requiremen...

If encouraged, as opposed to restrained, FOOM could happen days after AGI has the means to improve in the above dimensions. The hardware dimension would be limited at first by the human supply chain, and those limits would be gone when AGI runs its own supply chain - perhaps within a year after AGI.

Your logic is sound.

Tesla pundits predict big sales of Cybertruck wraps and accessories as customers seek the status of customization for an truck that many, many will end up having.

Tesla raised prices in order to reduce the time between order and delivery when customer demand got too hot. Raising prices also improved profit margins during a time of rapid capital spending. Now customer demand has cooled with the international economic slowdown, and lowering prices increases the time between order and delivery as more customers are enticed to order.

Intel and AMD both need to make LLM inference tools plug n play with their GPUs.

Otherwise Nvidia is the default choice when selecting a GPU for local LLM inference.

Aurora is working with OpenAI and Azure which is a very positive step in this direction.

On the other hand, Musk's other female CEO is wildly successful.

Witness Gwynne Shotwell at SpaceX. She dominates commercial space launch sales and the Starlink ISP constellation .

Twitter is much better technically in the past year because of Elon. Now sales and paid services will boom too.