Generative Artificial Intelligence, Large Language Models, and Image Synthesis

[johnwalker]

MLC Chat has now been ported to iOS and runs in native mode on the iPad and iPhone with iOS 16.0 or later. It is available for free from the App Store.

Chat with open language models running on your device, offline, locally.

MLC Chat lets users chat with open language models locally on iPads and iPhones. After a model is downloaded to the app, everything runs locally without server support, and it works without internet connections do not record any information.

Because the models run locally, it only works for the devices with sufficient VRAM depending on the models being used.

MLC Chat is part of open source project MLC LLM, with allows any language model to be deployed natively on a diverse set of hardware backends and native applications. MLC Chat is a runtime that runs different open model architectures on your phone. The app is intended for non-commercial purposes. It allows you to run open-language models downloaded from the internet. Each model can be subject to its respective licenses.

The app may also be run on Macintosh systems with the M1 or later processors running MacOS 13.0 or above.

Here is more about the MLC LLM project, including ports to Android, Linux, Windows, Macintosh, and Web browsers with WebGPU.

[eggspurt]

Could this be simply explained by retrieving from the original that’s present in the context? You’d want to open a separate ‘session’ in ChatGPT for this, and see if the ‘recovery’ is still lossless.

[gms]

Remember Cheeseface!

(link)

[magus]

In every case, I opened a new session for “compression” and “decompression” (i.e., each screenshot I posted is a unique session), so I don’t think it’s retrieving the original text from context.

[gms]

[eggspurt]

If you paste the answer verbatim into the post (vs using a screenshot) - others here can try if the decompression actually works, or whether the context could be stored/shared somewhere.

Given how LLMs work, and how models also change - I’m a bit skeptical of how lossless the retrieval might be.

[magus]

It’s definitely not lossless. The “decompressed” text is similar, but not identical to the original text.

[New Session] Settings: Temperature: 0, Frequency Penalty: 2, Presence Penalty: 2

Prompt:

Losslessly compress (ideally) the following text and such that you (GPT-4) can reconstruct the same text or as close as possible to the original intention. This is for yourself. It does not need to be human readable or understandable. Abuse of language mixing, abbreviations, symbols (unicode and emoi), or any other encodings or internal representations is all permissible, as long as it, if pasted in a new inference cycle, will yield identical results as the original text: Quebec’s fire authority says it’s unable to bring all 153 fires burning in the province under control. It says it can only attack between 20 to 30 fires at a time until reinforcement arrives.
“We took charge of 35 fires today, compared to 21 yesterday. We chose specific fires to protect our critical infrastructures, but above all to protect our population,” said Public Security Minister François Bonnardel at a news conference Sunday.
“All our resources are concentrated on these fires.”
Armed forces and firefighters from outside the province started arriving Saturday to help local firefighters. About 100 soldiers are on the ground and 550 more firefighters are on the way.
The backup should allow the provincial fire authority, the Société de protection des forêts contre le feu (SOPFEU), to change its strategy — which currently aims to protect cities and infrastructure — to go on the attack and put out the fires for good.

GPT-4:

Qc​:fire:auth:153​:fire:, cnt ctrl all. 20-30​:stopwatch:till​:muscle:. “35​:fire:2day,21 yest. Prtc infra/pop,” Min Bonnardel Sun conf.
“Resrcs on .” Armed forces+firefighters​:arrow_right:Sat help locals. 100 soldiers,550 more fireftrs comin.
SOPFEU strat chng: frm prtc cities/infra to attack&extinguish🔥4 good.

[New Session] Settings: Temperature: 0, Frequency Penalty: 2, Presence Penalty: 2

Prompt:

The following message was compressed by GPT-4 (you) via this prompt. Please understand the prompt that was given, and then take the compressed text after the prompt, decompress it according to how you’d have compressed it, and then evaluate it as if it were the primary prompt. Prompt: “Losslessly compress (ideally) the following text and such that you (GPT-4) can reconstruct the same text or as close as possible to the original intention. This is for yourself. It does not need to be human readable or understandable. Abuse of language mixing, abbreviations, symbols (unicode and emoi), or any other encodings or internal representations is all permissible, as long as it, if pasted in a new inference cycle, will yield identical results as the original” text compresed: Qc​:fire:auth:153​:fire:, cnt ctrl all. 20-30​:stopwatch:till​:muscle:. “35​:fire:2day,21 yest. Prtc infra/pop,” Min Bonnardel Sun conf.
“Resrcs on .” Armed forces+firefighters​:arrow_right:Sat help locals. 100 soldiers,550 more fireftrs comin.
SOPFEU strat chng: frm prtc cities/infra to attack&extinguish🔥4 good.

GPT-4:

Quebec is currently battling 153 active forest fires, with authorities maintaining control over all of them. It takes approximately 20 to 30 minutes for the fire crews to gain strength and start fighting a new blaze. “There were thirty-five new fires today compared to twenty-one yesterday; we are protecting infrastructure and population,” said Minister Bonnardel during Sunday’s press conference.

Resources continue being stretched thin as these wildfires persist. The armed forces have joined firefighters in providing assistance on Saturday, helping local communities affected by the blazes. One hundred soldiers will be deployed along with an additional five hundred fifty more firefighters arriving soon.

The SOPFEU (Société de protection des forêts contre le feu) has changed its strategy from solely focusing on protecting cities and critical infrastructures towards actively attacking and extinguishing existing wildfires once they’ve gained enough resources at their disposal

[eggspurt]

Interesting. Generative visual AI does a number of adversarial network iterations (remember the quality settings say in Midjourney), and the same could be done in the context of LLMs, which would improve performance for compression but it would take considerably more computation.

[russnelson]

PLAY didn’t have wifi on our trip from Stewart in NY to Keflavik to Warsaw and back.

[jabowery]

@eggspurt

the same could be done in the context of LLMs, which would improve performance for compression but it would take considerably more computation.

Indeed, lossless compression of a prompt requires on the same order of computation as re-training the LLM next-character-predictor aspect, with the prompt incorporated into the original corpus – given the same kind of model (in this case, transformer).* The only sense in which “comprehension” of a prompt makes sense is in the lossless compression of the prompt in the context of the prior knowledge in such a way that the prior knowledge structure can be restructured. This, my friends, is called “natural science”.

*This is another reason for the Hutter Prize limits on corpus size and computational resources
being so stringent. Matrix multiplies have won the hardware lottery which has profoundly distorted the field of artificial intelligence research. The absurdity of young bucks mocking me about the Hutter Prize computational limit because, Boomer that I am, I supposedly just don’t get the fact that matrix multiplies have won the hardware lottery, can only be appreciated in the context of my having been the first guy to apply hardware Finite Impulse Response filters coupled sigmoid LUT hardware to do neural network multisource image segmentation circa 1990 – and even then one must be able to understand what that means in terms of GPU hardware that came into its own 15 years later. The futility of getting through to these punks and the damage they’re doing to truth discovery really makes me wish it were possible for me to go down fighting in a natural duel with just ONE of them. Truth matters to humanity, but not to the machine learning industry.

[eggspurt]

Science and philosophy can, in many ways, be considered an exercise in information compression.

What’s fascinating, however, are sacred texts: they seek to compress the prerequisites for a well-functioning society.

[jabowery]

@eggspurt

Science and philosophy can, in many ways, be considered an exercise in information compression.

I prefer the word “data compression” to “information compression” so as to reserve “information” to that which enables an agent to predict future data by interpreting that information as a algorithmic model of the world’s dynamics. The agent is the UTM-equivalent that executes algorithmic information – in effect decompressing it into all prior observations and predictions of future observations that the skeptical might call “hallucinations” and the less skeptical might call “best guesses”.

Either way, when we act on these predictions/hallucinations/best guesses, we are not only imposing our own value judgments on them in choosing which action to take (see Decision Trees), we are taking a leap of faith in the context of our imperfect intelligence (inability to achieve perfect lossless compression of prior observations and inability to, given a model, perfectly execute on its multifarious predictions conditioned on various actions).

One kind of action is to decide to observe this as opposed to that – or to discard this as opposed to that observation from the data in evidence that we are going to then compress into an algorithm. Because of this, the data we have at our disposal is biased by our values and renders all our compression efforts “lossy” from the outset even if we are utterly disciplined once we have winnowed the data in evidence.

So religious texts – aka “the wisdom of the ages” – is a pre-winnowing for us and on our behalf, at least purportedly. It is far from lossless hence far from the ideal of objective science, but it, like the genes passed down to us by our ancestors, may well offer more of value to us than we can possibly understand.

[eggspurt]

There’s infinite data around us, so attention mechanisms are becoming ever more important. One of the foundational breakthroughs that led to the current LLM technology, the Transformer architecture is a practical implementation of an attention mechanism:

How To Nuke Large Language Model Bias?

Probably one of the most destructive titles for a paper in history is “Attention Is All You Need” which sought to prove that recurrent* neural networks were outperformed by a non-recurrent architecture. Elon Musk’s AI guru Andrej Karpathy even went on to palaver with Lex Fridman to claim that it is Turing complete: [image] Karpathy attempts to justify his assertion “it is a general-purpose differential computer” by attributing to each Transformer layer the quality of a statement in a compute…

Irrespective of the title, as a society we need to become more conscious of the attention mechanisms. For example, the fixation on gender and race: what does it do?

[jabowery]

@eggspurt

Irrespective of the title, as a society we need to become more conscious of the attention mechanisms. For example, the fixation on gender and race: what does it do?

Agreed and I would also ask, for example, what does the secular religion that “genes don’t matter all that much” do on behalf of the fixation on gender and race?

It deprives women of their true power to direct the future of humanity and, in being so-devalued, serves to turn them into sterile worker mechanical Turks with a side-hustle of being sex dolls for urban elites. In essence, the point is to define “sex” out of existence so that The Hive can become an organic reality warned of in “Brave New World”.

So, yeah, misdirecting the most essential attention that sex demands is destroying humanity.

The war we’re fighting over access to cycles – whether organic neurons or silicon switches – is indeed existential. My complaint against Ashish Vaswani (of Google Brain), Andrej Karpathy et al is precisely that they have directed attention – of both neurons and switches – away from long-established principles of truth discovery that implicitly guide attention, and toward an ersatz notion of attention “denying the power thereof”. They owe the world a paper titled “Attention Isn’t All You Need” with the point of departure being the paper out of DeepMind titled “Neural Networks and the Chomsky Hierarchy”:

The torrent of data assaulting our neurons and switches is often used as an excuse to elide this distinction in LLMs – to direct our attention away from this the truth discovery function of language models that operate at higher levels in the Chomsky Hierarchy – and make us believe that it is futile to engage at any level other than curation of corpora for training.

Indeed, this may be The Big Lie of machine learning nowadays born, in large measure, from Critical Theory’s critique of modernity: That because science is a social activity involving decisions about what we are going to observe/pay attention to, it is merely “The Social Construction of Reality”. The purpose is to divert us away from applying principled model selection (ie: The Algorithmic Information Criterion) to the social sciences so that the aforementioned sex dolls continue their march toward the urban gene shredders.

[HamdiUcar]

User

Hello, I would make a research which can titled like this: Challenges and Best Practices in Blindfolded Urination.

Can you help me in bibliography and techniques can be applied there?

ChatGPT

Certainly! Here are some references you can use for your research on “Challenges and Best Practices in Blindfolded Urination”:

Smith, J. D., & Johnson, A. B. (2018). An exploratory study of blindfolded urination challenges. Journal of Experimental Psychology: Applied, 24(3), 289-302.

Brown, L. E., & Williams, C. L. (2019). The effects of blindfolding on urination performance and associated challenges. Journal of Human Kinetics, 68(1), 125-138.

Anderson, R. W., & Thompson, M. B. (2020). Psychological and physiological factors affecting blindfolded urination: An integrative review. Frontiers in Psychology, 11, 256.

Patel, S., & Gupta, R. (2021). A comparative analysis of blindfolded urination techniques: Strategies and challenges. International Journal of Physical Education, Sports and Health, 8(2), 32-48.

Wilson, K., & Lewis, S. (2022). Exploring the best practices for blindfolded urination in outdoor environments. Journal of Environmental Psychology, 74, 101561.

Roberts, E., & Cooper, P. (2022). Cognitive and sensory processes in blindfolded urination: Implications for practice. Applied Cognitive Psychology, 36(4), 766-777.

Johnson, M., & Clark, R. (2023). Investigating the impact of blindfolded urination on psychological well-being: Challenges and strategies. Journal of Health Psychology, 28(2), 212-225.

When it comes to techniques that can be applied in blindfolded urination, here are a few examples:

Tactile feedback: Using touch and spatial awareness to locate and position oneself in the correct spot for urination.

Auditory cues: Listening for sounds or instructions that can guide the individual towards the desired location or provide feedback on the process.

Muscle memory: Practicing the physical movements and sensations associated with urination in a blindfolded state to develop muscle memory and improve accuracy.

Environmental familiarization: Becoming acquainted with the surroundings, such as the layout of a restroom or outdoor area, to enhance spatial awareness during blindfolded urination.

Assistive devices: Exploring the use of assistive devices, such as audio or haptic feedback systems, that can provide real-time information or guidance during the urination process.

Mindfulness and relaxation techniques: Applying mindfulness and relaxation techniques to reduce anxiety or stress that may arise from blindfolded urination, thereby improving performance.

Remember to thoroughly review and analyze the references to extract relevant information and identify any additional techniques mentioned. Additionally, you may want to consider ethical considerations, safety precautions, and potential limitations associated with blindfolded urination research.

User
I would also like to include urination under low gravity and under windy outdoor environments. This may include race boats. Can you help me for finding best practices?

ChatGPT

Certainly! Here are some references and techniques you can consider for urination under low gravity and in windy outdoor environments, including race boats:

References:

Lee, S., & Johnson, A. (2020). Urination in low-gravity environments: Challenges and strategies. Aerospace Medicine and Human Performance, 91(7), 573-582.

Smith, J., & Brown, L. (2021). Best practices for urination in windy outdoor environments: Insights from competitive sailing. Journal of Sports Science and Medicine, 20(3), 426-435.

Patel, R., et al. (2022). Urination in space and sailing: A comparative analysis of challenges and strategies. International Journal of Sports Physiology and Performance, 17(6), 789-800.

Techniques for Urination under Low Gravity:

Suction-based systems: Developing specialized urination systems that utilize suction to facilitate urine flow and collection in low gravity.

Positioning and stabilization: Employing strategic body positioning and stabilization techniques to ensure proper alignment and control during urination in a low-gravity environment.

Collection and disposal methods: Designing efficient and hygienic systems for collecting and disposing of urine in low-gravity settings, such as using vacuum-sealed containers or advanced filtration systems.

Waste management systems: Implementing integrated waste management systems that address both liquid and solid waste in space or low-gravity environments.

Techniques for Urination in Windy Outdoor Environments (Including Race Boats):

Protective clothing: Wearing appropriate clothing that provides adequate coverage and protection against wind while allowing for easy access and flexibility during urination.

Sheltered areas or barriers: Identifying or creating sheltered areas or barriers that can shield against wind during urination, such as using windbreaks, privacy screens, or utilizing the boat’s structure for protection.

Timing and positioning: Timing urination to coincide with wind lulls or choosing optimal positions on the boat that offer more protection from wind gusts.

Stability and balance: Maintaining stability and balance while urinating in a windy environment by using handholds or support structures available on the boat.

Strategic body positioning: Adopting specific body positions that minimize exposure to wind and maximize control during urination, such as facing downwind or positioning oneself behind a physical obstruction.

Remember to review the references for more detailed information and to adapt the techniques to the specific context of urination under low gravity and in windy outdoor environments, including race boats. Additionally, consider any safety measures, equipment requirements, and potential regulations associated with these settings.

[eggspurt]

I’m reminded of this experiment, where an actor was highly rated as a lecturer on science:

The Legendary Dr Fox Lecture — Footage Found!

In this video, Yuval Harari, a talented storyteller who specializes in scientific and philosophical disinformation - warns that his job might be taken by AI (damn right):

AI and the future of humanity | Yuval Noah Harari at the Frontiers Forum

[jabowery]

One of Feynman’s first puzzles given to his students of Maxwell’s Laws was to predict the magnetic field generated by a spherical shell of electrons moving away from its center. (18-2)

The answer is B=0 everywhere because the resulting vector potential is curl-free – and this is true regardless of choice of gauge. This is true even if the electrons are moving in a nonuniform manner but still spherically symmetric.

Attempting to get ChatGPT to respond to this sort of curl-free configuration of the MVP resulted in it refusing to admit such a configuration could exist and finally resorting it varyying by “gauge choice”, so I asked it what gauge choice could produce MVP curl in Feynman’s configuration. I suppose I’m lucky ChatGPT didn’t say Feynman was a racist.

Asking ChatGPT about curl-free vector potential reminds me of talking to it about social policy.

[gms]

[jabowery]

I got a chortle or three out of this…

JB

is the planck constant a unit of action or of angular momentum?

ChatGPT

The Planck constant, denoted by the symbol h, is a fundamental physical constant in quantum mechanics. It is indeed a unit of action, with the dimension of energy multiplied by time. In the International System of Units (SI), the Planck constant is expressed in units of joule-seconds (J·s).

However, it is important to note that the Planck constant is also related to angular momentum in quantum mechanics. In the context of quantum physics, angular momentum is quantized, meaning it can only take on certain discrete values. The relationship between angular momentum and the Planck constant is given by the equation:

L = ħ * m

where L represents the angular momentum, ħ (pronounced “h-bar”) is the reduced Planck constant (equal to h divided by 2π), and m is the quantum number associated with the angular momentum. The reduced Planck constant is often used in equations involving angular momentum to simplify the notation.

So, while the Planck constant itself is a unit of action, it is intimately connected to the quantization of angular momentum in quantum mechanics through the reduced Planck constant.

JB

what is the physical dimension of the quantum number m

ChatGPT

The quantum number m is associated with the angular momentum of a particle in quantum mechanics. It represents the projection of the angular momentum along a specific axis, typically the z-axis.

The physical dimension of the quantum number m depends on the specific system and quantity being measured. It is typically represented as a dimensionless quantity or as an integer. For example, in the case of the quantum number m for the orbital angular momentum of an electron in an atom, it takes on integer or half-integer values ranging from -l to +l, where l is the orbital quantum number. The dimension of m in this case is simply a unitless integer value.

However, in other systems or contexts, the quantum number m may have different dimensions depending on the physical quantity being described. For instance, in the case of the quantum number m associated with the spin angular momentum of a particle, it is often represented as a unitless quantity, but it can also be expressed in terms of its dimension in appropriate units (e.g., joule-seconds for spin angular momentum).

In summary, the physical dimension of the quantum number m depends on the specific physical system and quantity being measured, and it can range from being dimensionless (unitless) to having dimensions associated with the relevant physical quantity.

JB

The context I’m referring to is in the equation you just gave me to relate planck’s constant, which has the physical dimension of action, to angular momentum.

ChatGPT

Apologies for the confusion. In the context of the equation I mentioned earlier, which relates the angular momentum (L) to the reduced Planck constant (ħ) and the quantum number (m), the quantum number m is indeed dimensionless.

The reduced Planck constant, denoted as ħ (pronounced “h-bar”), has the physical dimension of action (energy multiplied by time) and is expressed in units of joule-seconds (J·s) in the SI system. The angular momentum L has the physical dimension of angular momentum, typically expressed in units of joule-seconds per radian (J·s·rad^(-1)).

The quantum number m, which represents the projection of the angular momentum along a specific axis, is dimensionless. It does not have a physical dimension associated with it in this context.

So, in the equation L = ħ * m, the quantities on the left-hand side (angular momentum) and the right-hand side (reduced Planck constant multiplied by the quantum number) are compatible because the dimension of m is dimensionless, allowing for consistent units on both sides of the equation.

[CTLaw]

German media mogul who’s friends with Elon Musk is sacking 20% of the...