The A.I Megathread (LLM , GPT , Development)

[bnew]

---

https://huggingface.co/ehartford/dolphin-2_2-yi-34b

Dolphin 2.2 https://erichartford.com/dolphin

Dolphin-2.2-Yi-34b's training was sponsored by a16z.

This model is based on Yi, and is subject to Yi license.

I used the llama compatible chargoddard/Yi-34B-Llama as the base model.

Trained with 16k context. You can load it as follows:

from transformers import LlamaForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("ehartford/dolphin-2.2-yi-34b", trust_remote_code=True)
model = LlamaForCausalLM.from_pretrained("ehartford/dolphin-2.2-yi-34b")
New in 2.2 is conversation and empathy. With an infusion of curated Samantha and WizardLM DNA, Dolphin can now give you personal advice and will care about your feelings, and with extra training in long multi-turn conversation.

This model is uncensored. I have filtered the dataset to remove alignment and bias. This makes the model more compliant. You are advised to implement your own alignment layer before exposing the model as a service. It will be highly compliant to any requests, even unethical ones. Please read my blog post about uncensored models. https://erichartford.com/uncensored-models You are responsible for any content you create using this model. Enjoy responsibly.

Dataset​

This dataset is Dolphin, an open-source implementation of Microsoft's Orca

I modified the dataset for uncensoring, deduping, cleaning, and quality.

I added Jon Durbin's excellent Airoboros dataset to increase creativity.

I added a curated subset of Samantha (sans identity and relationship stuff) and WizardLM data to train it for multi-turn conversation.

Training​

It took 3 days to train 3 epochs on 4x A100s using qLoRA and Axolotl

https://huggingface.co/ehartford/dolphin-2_2-yi-34b

https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-GGUF

 

[bnew]

https://archive.is/NWDiZ

---

 

[Matt504]

YG's 'Meet the Flockers' Returns to Streaming With Anti-Asian Lyrics Removed

YG's "Meet the Flockers" has returned to streaming services after disappearing this week following controversy over its anti-Asian lyrics.

www.rollingstone.com

 

[bnew]

GraphCast: AI model for faster and more accurate global weather forecasting

Our state-of-the-art model delivers 10-day weather predictions at unprecedented accuracy in under one minute

deepmind.google

GraphCast: AI model for faster and more accurate global weather forecasting​

Published14 NOVEMBER 2023Authors

Remi Lam on behalf of the GraphCast team

Our state-of-the-art model delivers 10-day weather predictions at unprecedented accuracy in under one minute

The weather affects us all, in ways big and small. It can dictate how we dress in the morning, provide us with green energy and, in the worst cases, create storms that can devastate communities. In a world of increasingly extreme weather, fast and accurate forecasts have never been more important.

In a paper published in Science, we introduce GraphCast, a state-of-the-art AI model able to make medium-range weather forecasts with unprecedented accuracy. GraphCast predicts weather conditions up to 10 days in advance more accurately and much faster than the industry gold-standard weather simulation system – the High Resolution Forecast (HRES), produced by the European Centre for Medium-Range Weather Forecasts (ECMWF).

GraphCast can also offer earlier warnings of extreme weather events. It can predict the tracks of cyclones with great accuracy further into the future, identifies atmospheric rivers associated with flood risk, and predicts the onset of extreme temperatures. This ability has the potential to save lives through greater preparedness.

GraphCast takes a significant step forward in AI for weather prediction, offering more accurate and efficient forecasts, and opening paths to support decision-making critical to the needs of our industries and societies. And, by open sourcing the model code for GraphCast, we are enabling scientists and forecasters around the world to benefit billions of people in their everyday lives. GraphCast is already being used by weather agencies, including ECMWF, which is running a live experiment of our model’s forecasts on its website.

Watch



A selection of GraphCast’s predictions rolling across 10 days showing specific humidity at 700 hectopascals (about 3 km above surface), surface temperature, and surface wind speed.

The challenge of global weather forecasting​

Weather prediction is one of the oldest and most challenging–scientific endeavours. Medium range predictions are important to support key decision-making across sectors, from renewable energy to event logistics, but are difficult to do accurately and efficiently.

Forecasts typically rely on Numerical Weather Prediction (NWP), which begins with carefully defined physics equations, which are then translated into computer algorithms run on supercomputers. While this traditional approach has been a triumph of science and engineering, designing the equations and algorithms is time-consuming and requires deep expertise, as well as costly compute resources to make accurate predictions.

Deep learning offers a different approach: using data instead of physical equations to create a weather forecast system. GraphCast is trained on decades of historical weather data to learn a model of the cause and effect relationships that govern how Earth’s weather evolves, from the present into the future.

Crucially, GraphCast and traditional approaches go hand-in-hand: we trained GraphCast on four decades of weather reanalysis data, from the ECMWF’s ERA5 dataset. This trove is based on historical weather observations such as satellite images, radar, and weather stations using a traditional NWP to ‘fill in the blanks’ where the observations are incomplete, to reconstruct a rich record of global historical weather.

GraphCast: An AI model for weather prediction​

GraphCast is a weather forecasting system based on machine learning and Graph Neural Networks (GNNs), which are a particularly useful architecture for processing spatially structured data.

GraphCast makes forecasts at the high resolution of 0.25 degrees longitude/latitude (28km x 28km at the equator). That’s more than a million grid points covering the entire Earth’s surface. At each grid point the model predicts five Earth-surface variables – including temperature, wind speed and direction, and mean sea-level pressure – and six atmospheric variables at each of 37 levels of altitude, including specific humidity, wind speed and direction, and temperature.

While GraphCast’s training was computationally intensive, the resulting forecasting model is highly efficient. Making 10-day forecasts with GraphCast takes less than a minute on a single Google TPU v4 machine. For comparison, a 10-day forecast using a conventional approach, such as HRES, can take hours of computation in a supercomputer with hundreds of machines.

In a comprehensive performance evaluation against the gold-standard deterministic system, HRES, GraphCast provided more accurate predictions on more than 90% of 1380 test variables and forecast lead times (see our Science paper for details). When we limited the evaluation to the troposphere, the 6-20 kilometer high region of the atmosphere nearest to Earth’s surface where accurate forecasting is most important, our model outperformed HRES on 99.7% of the test variables for future weather.

For inputs, GraphCast requires just two sets of data: the state of the weather 6 hours ago, and the current state of the weather. The model then predicts the weather 6 hours in the future. This process can then be rolled forward in 6-hour increments to provide state-of-the-art forecasts up to 10 days in advance.

Better warnings for extreme weather events​

Our analyses revealed that GraphCast can also identify severe weather events earlier than traditional forecasting models, despite not having been trained to look for them. This is a prime example of how GraphCast could help with preparedness to save lives and reduce the impact of storms and extreme weather on communities.

By applying a simple cyclone tracker directly onto GraphCast forecasts, we could predict cyclone movement more accurately than the HRES model. In September, a live version of our publicly available GraphCast model, deployed on the ECMWF website, accurately predicted about nine days in advance that Hurricane Lee would make landfall in Nova Scotia. By contrast, traditional forecasts had greater variability in where and when landfall would occur, and only locked in on Nova Scotia about six days in advance.

GraphCast can also characterize atmospheric rivers – narrow regions of the atmosphere that transfer most of the water vapour outside of the tropics. The intensity of an atmospheric river can indicate whether it will bring beneficial rain or a flood-inducing deluge. GraphCast forecasts can help characterize atmospheric rivers, which could help planning emergency responses together with AI models to forecast floods.

Finally, predicting extreme temperatures is of growing importance in our warming world. GraphCast can characterize when the heat is set to rise above the historical top temperatures for any given location on Earth. This is particularly useful in anticipating heat waves, disruptive and dangerous events that are becoming increasingly common.

Severe-event prediction - how GraphCast and HRES compare.


Left: Cyclone tracking performances. As the lead time for predicting cyclone movements grows, GraphCast maintains greater accuracy than HRES.


Right: Atmospheric river prediction. GraphCast’s prediction errors are markedly lower than HRES’s for the entirety of their 10-day predictions

The future of AI for weather​

GraphCast is now the most accurate 10-day global weather forecasting system in the world, and can predict extreme weather events further into the future than was previously possible. As the weather patterns evolve in a changing climate, GraphCast will evolve and improve as higher quality data becomes available.

To make AI-powered weather forecasting more accessible, we’ve open sourced our model’s code. ECMWF is already experimenting with GraphCast’s 10-day forecasts and we’re excited to see the possibilities it unlocks for researchers – from tailoring the model for particular weather phenomena to optimizing it for different parts of the world.

GraphCast joins other state-of-the-art weather prediction systems from Google DeepMind and Google Research, including a regional Nowcasting model that produces forecasts up to 90 minutes ahead, and MetNet-3, a regional weather forecasting model already in operation across the US and Europe that produces more accurate 24-hour forecasts than any other system.

Pioneering the use of AI in weather forecasting will benefit billions of people in their everyday lives. But our wider research is not just about anticipating weather – it’s about understanding the broader patterns of our climate. By developing new tools and accelerating research, we hope AI can empower the global community to tackle our greatest environmental challenges.

---

 

[bnew]

Spoiler: thread continued

 

[bnew]

https://archive.is/dyU4M

 

[bnew]

https://archive.is/UbyJA

Spoiler: thread continued

https://latent-consistency-models.github.io/

---

https://arxiv.org/abs/2311.05556

LCM-LoRA: A Universal Stable-Diffusion Acceleration Module​

Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, Hang Zhao

Latent Consistency Models (LCMs) have achieved impressive performance in accelerating text-to-image generative tasks, producing high-quality images with minimal inference steps. LCMs are distilled from pre-trained latent diffusion models (LDMs), requiring only ~32 A100 GPU training hours. This report further extends LCMs' potential in two aspects: First, by applying LoRA distillation to Stable-Diffusion models including SD-V1.5, SSD-1B, and SDXL, we have expanded LCM's scope to larger models with significantly less memory consumption, achieving superior image generation quality. Second, we identify the LoRA parameters obtained through LCM distillation as a universal Stable-Diffusion acceleration module, named LCM-LoRA. LCM-LoRA can be directly plugged into various Stable-Diffusion fine-tuned models or LoRAs without training, thus representing a universally applicable accelerator for diverse image generation tasks. Compared with previous numerical PF-ODE solvers such as DDIM, DPM-Solver, LCM-LoRA can be viewed as a plug-in neural PF-ODE solver that possesses strong generalization abilities. Project page: this https URL.

Comments:	Technical Report
Subjects:	Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
Cite as:	arXiv:2311.05556 [cs.CV]
	(or arXiv:2311.05556v1 [cs.CV] for this version)
	https://doi.org/10.48550/arXiv.2311.05556 Focus to learn more

https://arxiv.org/pdf/2311.05556.pdf

 

[bnew]

Training of 1-Trillion Parameter Scientific AI Begins

A US national lab has started training a massive AI brain that could ultimately become the must-have computing resource for scientific researchers. Argonne National Laboratory (ANL) is creating a generative […]

www.hpcwire.com

Training of 1-Trillion Parameter Scientific AI Begins​

By Agam Shah

November 13, 2023

A US national lab has started training a massive AI brain that could ultimately become the must-have computing resource for scientific researchers.

Argonne National Laboratory (ANL) is creating a generative AI model called AuroraGPT and is pouring a giant mass of scientific information into creating the brain.

The model is being trained on its Aurora supercomputer, which delivers more than an half an exaflop performance at ANL. The system has Intel’s Ponte Vecchio GPUs, which provide the main computing power.

Intel and ANL are partnering with other labs in the US and worldwide to make scientific AI a reality.
“It combines all the text, codes, specific scientific results, papers, into the model that science can use to speed up research,” said Ogi Brkic, vice president and general manager for data center and HPC solutions, in a press briefing.

Brkic called the model “ScienceGPT,” indicating it will have a chatbot interface, and researchers can submit questions and get responses.

Chatbots could help in a wide range of scientific research, including biology, cancer research, and climate change.

Example ChatGPT v3.5 Question and Answer

Training a model with complex data can take time and massive computing resources. ANL and Intel are in the early stages of testing the hardware before putting the model into full training mode.

While it will operate like ChatGPT, it is unclear if the generative model will be multimodal or whether it will generate images and videos. Inference will also be a big part of the system as scientists seek answers from the chatbot and feed more information into the model.

Training AuroraGPT has just started and could take months to complete. The training is currently limited to 256 nodes, which will then be scaled to all of the nodes — about 10,000 — of the Aurora supercomputer.

OpenAI has not shared details on how long it took to train GPT-4, which takes place on Nvidia GPUs. In May, Google said it was training its upcoming large-language model called Gemini, which is likely happening on its TPUs.

The biggest challenge in training large language models is the memory requirements, and in most cases, the training needs to be sliced down to smaller bits across a wide range of GPUs. The AuroraGPT is enabled by Microsoft’s Megatron/DeepSpeed, which does exactly that and ensures the training is happening in parallel.

Intel and ANL are testing the 1-trillion parameter model training on a string of 64 Aurora nodes.
“The number of nodes is lower than we typically see on these large language models… because [of the] unique Aurora design,” Brkic said.

Intel has worked with Microsoft on fine-tuning the software and hardware, so the training can scale to all nodes. The goal is to extend this to the entire system of 10,000 plus nodes.

Intel also hopes to achieve linear scaling so the performance increases as the number of nodes increases.

Brkic said its Ponte Vecchio GPUs outperformed Nvidia’s A100 GPUs in another Argonne supercomputer called Theta, which has a peak performance of 11.7 petaflops.

 

[bnew]

https://github.com/lightly-ai/labelformat

About​

A tool for converting computer vision label formats.

Labelformat - Label Conversion, Simplified​

An open-source tool to seamlessly convert between popular computer vision label formats.

Why Labelformat​

Popular label formats are sparsely documented and store different information. Understanding them and dealing with the differences is tedious and time-consuming. Labelformat aims to solve this pain.

Supported Tasks and Formats​

• object-detection
  • COCO
  • KITTI
  • Lightly
  • PascalVOC
  • YOLOv5
  • YOLOv6
  • YOLOv7
  • YOLOv8
  • Labelbox (input only)
• instance-segmentation
  • COCO
  • YOLOv8

Features​

• Support for common dataset label formats (more coming soon)
• Support for common tool formats (more coming soon)
• Minimal dependencies, targets python 3.7 or higher
• Memory concious - datasets are processed file-by-file instead of loading everything in memory (when possible)
• Typed
• Tested with round trip tests to ensure consistency
• MIT license

Supported Platforms​

This package is compatible with the following platforms:

• Windows
• macOS
• Linux

 

[bnew]

YouTube is going to start cracking down on AI clones of musicians

Music labels are getting a way to remove AI content.

www.theverge.com

YouTube is going to start cracking down on AI clones of musicians​

/

YouTube will give music labels a way to take down content that ‘mimics an artist’s unique singing or rapping voice.’ Creators will also be required to label AI-generated content beginning next year.​

By Mia Sato and Nilay Patel

Nov 14, 2023, 6:33 AM EST|9 Comments / 9 New

Share this story​

• 
  
  If you buy something from a Verge link, Vox Media may earn a commission. See our ethics statement.
  
  
  
  Photo by Prince Williams/Wireimage
  
  YouTube will have two sets of content guidelines for AI-generated deepfakes: a very strict set of rules to protect the platform’s music industry partners, and another, looser set for everyone else.
  
  That’s the explicit distinction laid out today in a company blog post, which goes through the platform’s early thinking about moderating AI-generated content. The basics are fairly simple: YouTube will require creators to begin labeling “realistic” AI-generated content when they’re uploading videos, and that the disclosure requirement is especially important for topics like elections or ongoing conflicts.
  
  The labels will appear in video descriptions, and on top of the videos themselves for sensitive material. There is no specific definition of what YouTube thinks “realistic” means yet; YouTube spokesperson Jack Malon tells us that the company will provide more detailed guidance with examples when the disclosure requirement rolls out next year.
  
  
  
  
  YouTube will require creators to label “realistic” AI-generated content. Image: YouTube
  
  YouTube says the penalties for not labeling AI-generated content accurately will vary, but could include takedowns and demonetization. But it’s not clear how YouTube will know if an unlabeled video was actually generated by AI — YouTube’s Malon says the platform is “investing in the tools to help us detect and accurately determine if creators have fulfilled their disclosure requirements when it comes to synthetic or altered content,” but those tools don’t exist yet and the ones that do have notoriously poor track records.
  
  From there, it gets more complicated — vastly more complicated. YouTube will allow people to request removal of videos that “simulate an identifiable individual, including their face or voice” using the existing privacy request form. So if you get deepfaked, there’s a process to follow that may result in that video coming down — but the company says it will “evaluate a variety of factors when evaluating these requests,” including whether the content is parody or satire and whether the individual is a public official or “well-known individual.”
  
  
  
  
  AI labels will appear on videos or in descriptions, depending on the content. Image: YouTube
  
  If that sounds vaguely familiar, it’s because those are the same sorts of analyses courts do: parody and satire is an important element of the fair use defense in copyright infringement cases, and assessing whether someone is a public figure is an important part of defamation law. But since there’s no specific federal law regulating AI deepfakes, YouTube is making up its own rules to get ahead of the curve — rules which the platform will be able to enforce any way it wants, with no particular transparency or consistency required, and which will sit right alongside the normal creator dramas around fair use and copyright law.
  
  It is going to be wildly complicated — there’s no definition of “parody and satire” for deepfake videos yet, but Malon again said there would be guidance and examples when the policy rolls out next year.
  
  Making things even more complex, there will be no exceptions for things like parody and satire when it comes to AI-generated music content from YouTube’s partners “that mimics an artist’s unique singing or rapping voice,” meaning Frank Sinatra singing The Killers’ Mr. Brightside is likely in for an uphill battle if Universal Music Group decides it doesn’t like it.
  
  There are entire channels dedicated to churning out AI covers by artists living and dead, and under YouTube’s new rules, most would be subject to takedowns by the labels. The only exception YouTube offers in its blog post is if the content is “the subject of news reporting, analysis or critique of the synthetic vocals” — another echo of a standard fair use defense without any specific guidelines yet. YouTube has long been a generally hostile environment for music analysis and critique because of overzealous copyright enforcement, so we’ll have to see if the labels can show any restraint at all — and if YouTube actually pushes back.
  
  This special protection for singing and rapping voices won’t be a part of YouTube’s automated Content ID system when it rolls out next year; Malon tells us that “music removal requests will be made via a form” that partner labels will have to fill out manually. And the platform isn’t going to penalize creators who trip over these blurred lines, at least not in these early days — Malon says “content removed for either a privacy request or a synthetic vocals request will not result in penalties for the uploader.”
  
  YouTube is walking quite a tightrope here, as there is no established legal framework for copyright law in the generative AI era — there’s no specific law or court case that says it’s illegal to train an AI system to sing in Taylor Swift’s voice. But YouTube is also existentially dependent on the music industry — it needs licenses for all the music that floods the platform daily, and especially to compete with TikTok, which has emerged as the most powerful music discovery tool on the internet. There’s a reason YouTube and Universal Music noisily announced a deal to work on AI soon after Ghostwriter99 posted “Heart on my Sleeve” with the AI-generated voices of Drake and The Weeknd — YouTube has to keep these partners happy, even if that means literally taking the law into its own hands.
  
  At the same time, YouTube parent company Google is pushing ahead on scraping the entire internet to power its own AI ambitions — resulting in a company that is at once writing special rules for the music industry while telling everyone else that their work will be taken for free. The tension is only going to keep building — and at some point, someone is going to ask Google why the music industry is so special.

 

[bnew]

https://archive.is/pHc08

Shared November 13, 2023

Text-to-music generation models are now capable of generating high-quality music audio in broad styles. However, text control is primarily suitable for the manipulation of global musical attributes like genre, mood, and tempo, and is less suitable for precise control over time-varying attributes such as the positions of beats in time or the changing dynamics of the music. We propose Music ControlNet, a diffusion-based music generation model that offers multiple precise, time-varying controls over generated audio. To imbue text-to-music models with time- varying control, we propose an approach analogous to pixel-wise control of the image-domain ControlNet method. Specifically, we extract controls from training audio yielding paired data, and fine-tune a diffusion-based conditional generative model over audio spectrograms given melody, dynamics, and rhythm controls. While the image-domain Uni-ControlNet method already allows generation with any subset of controls, we devise a new strategy to allow creators to input controls that are only partially specified in time. We evaluate both on controls extracted from audio and controls we expect creators to provide, demonstrating that we can generate realistic music that corresponds to control inputs in both settings. While few comparable music generation models exist, we benchmark against MusicGen, a recent model that accepts text and melody input, and show that our model generates music that is 49% more faithful to input melodies despite having 35x fewer parameters, training on 11x less data, and enabling two additional forms of time-varying control. Sound examples can be found at MusicControlNet.github.io/web/.

 

[bnew]

Paper page - The Impact of Large Language Models on Scientific Discovery: a Preliminary Study using GPT-4

Join the discussion on this paper page

huggingface.co

The Impact of Large Language Models on Scientific Discovery: a Preliminary Study using GPT-4​

Published on Nov 13
·Featured in Daily Papers on Nov 13
Authors:
Microsoft Research AI4Science,
Microsoft Azure Quantum

Abstract​

In recent years, groundbreaking advancements in natural language processing have culminated in the emergence of powerful large language models (LLMs), which have showcased remarkable capabilities across a vast array of domains, including the understanding, generation, and translation of natural language, and even tasks that extend beyond language processing. In this report, we delve into the performance of LLMs within the context of scientific discovery, focusing on GPT-4, the state-of-the-art language model. Our investigation spans a diverse range of scientific areas encompassing drug discovery, biology, computational chemistry (density functional theory (DFT) and molecular dynamics (MD)), materials design, and partial differential equations (PDE). Evaluating GPT-4 on scientific tasks is crucial for uncovering its potential across various research domains, validating its domain-specific expertise, accelerating scientific progress, optimizing resource allocation, guiding future model development, and fostering interdisciplinary research. Our exploration methodology primarily consists of expert-driven case assessments, which offer qualitative insights into the model's comprehension of intricate scientific concepts and relationships, and occasionally benchmark testing, which quantitatively evaluates the model's capacity to solve well-defined domain-specific problems. Our preliminary exploration indicates that GPT-4 exhibits promising potential for a variety of scientific applications, demonstrating its aptitude for handling complex problem-solving and knowledge integration tasks. Broadly speaking, we evaluate GPT-4's knowledge base, scientific understanding, scientific numerical calculation abilities, and various scientific prediction capabilities.

https://arxiv.org/pdf/2311.07361.pdf

Our observations​

GPT-4 demonstrates considerable potential in various scientific domains, including drug discovery, biology, computational chemistry, materials design, and PDEs. Its capabilities span a wide range of tasks and it exhibits an impressive understanding of key concepts in each domain.

The output of GPT-4 depends on several variables such as the model version, system messages, and hyperparameters like the decoding temperature. Thus, one might observe different responses for the same cases examined in this report. For the majority of this report, we primarily utilized GPT-4 version 0314, with a few cases employing version 0613.

Azure OpenAI Service – Advanced Language Models | Microsoft Azure

Azure OpenAI Service offers industry-leading coding and language AI models that you can fine-tune to your specific needs for a variety of use cases.

azure.microsoft.com

The qualitative approach used in this report mainly refers to case studies. It is related to but not identical to qualitative
methods in social science research.


In drug discovery, GPT-4 shows a comprehensive grasp of the field, enabling it to provide useful insights and suggestions across a wide range of tasks. It is helpful in predicting drug-target binding affinity, molecular properties, and retrosynthesis routes. It also has the potential to generate novel molecules with desired properties, which can lead to the discovery of new drug candidates with the potential to address unmet medical needs. However, it is important to be aware of GPT-4’s limitations, such as challenges in processing SMILES sequences and limitations in quantitative tasks.

In the field of biology, GPT-4 exhibits substantial potential in understanding and processing complex biological language, executing bioinformatics tasks, and serving as a scientific assistant for biology design. Its extensive grasp of biological concepts and its ability to perform various tasks, such as processing specialized files, predicting signaling peptides, and reasoning about plausible mechanisms from observations, benefit it to be a valuable tool in advancing biological research. However, GPT-4 has limitations when it comes to processing biological sequences (e.g., DNA and FASTA sequences) and its performance on tasks related to under-studied entities.

In computational chemistry, GPT-4 demonstrates remarkable potential across various subdomains, including electronic structure methods and molecular dynamics simulations. It is able to retrieve information, suggest design principles, recommend suitable computational methods and software packages, generate code for various programming languages, and propose further research directions or potential extensions. However, GPT-4 may struggle with generating accurate atomic coordinates of complex molecules, handling raw atomic coordinates, and performing precise calculations.

In materials design, GPT-4 shows promise in aiding materials design tasks by retrieving information, suggesting design principles, generating novel and feasible chemical compositions, recommending analytical and numerical methods, and generating code for different programming languages. However, it encounters challenges in representing and proposing more complex structures, e.g., organic polymers and MOFs, generating accurate atomic coordinates, and providing precise quantitative predictions.

In the realm of PDEs, GPT-4 exhibits its ability to understand the fundamental concepts, discern relationships between concepts, and provide accurate proof approaches. It is able to recommend appropriate analytical and numerical methods for addressing various types of PDEs and generate code in different programming languages to numerically solve PDEs. However, GPT-4’s proficiency in mathematical theorem proving still has room for growth, and its capacity for independently discovering and validating novel mathematical theories remains limited in scope.

In summary, GPT-4 exhibits both significant potential and certain limitations for scientific discovery.

To better leverage GPT-4, researchers should be cautious and verify the model’s outputs, experiment with different prompts, and combine its capabilities with dedicated AI models or computational tools to ensure reliable conclusions and optimal performance in their respective research domains:

• Interpretability and Trust: It is crucial to maintain a healthy skepticism when interpreting GPT-4’s output. Researchers should always critically assess the generated results and cross-check them with existing knowledge or expert opinions to ensure the validity of the conclusions.

• Iterative Questioning and Refinement: GPT-4’s performance can be improved by asking questions in an iterative manner or providing additional context. If the initial response from GPT-4 is not satisfactory, researchers can refine their questions or provide more information to guide the model toward a more accurate and relevant answer.

• Combining GPT-4 with Domain-Specific Tools: In many cases, it may be beneficial to combine GPT-4’s capabilities with more specialized tools and models designed specifically for scientific discovery tasks, such as molecular docking software, or protein folding algorithms. This combination can help researchers leverage the strengths of both GPT-4 and domain-specific tools to achieve more reliable and accurate results. Although we do not extensively investigate the integration of LLMs and domain-specific tools/models in this report, a few examples are briefly discussed in Section 7.2.1.

 

[bnew]

Will Large Language Models End Programming?

Explore the advancements in LLM with OpenAI's GPT-4 Turbo, Copilot, etc. in transformative impact on programming and creative industries.

www.unite.ai

ARTIFICIAL INTELLIGENCE​

Will Large Language Models End Programming?​

Published
8 seconds ago
on
November 14, 2023

By
Aayush Mittal

Last week marked a significant milestone for OpenAI, as they unveiled GPT-4 Turbo at their OpenAI DevDay. A standout feature of GPT-4 Turbo is its expanded context window of 128,000, a substantial leap from GPT-4's 8,000. This enhancement enables the processing of text 16 times greater than its predecessor, equivalent to around 300 pages of text.

This advancement ties into another significant development: the potential impact on the landscape of SaaS startups.

OpenAI's ChatGPT Enterprise, with its advanced features, poses a challenge to many SaaS startups. These companies, which have been offering products and services around ChatGPT or its APIs, now face competition from a tool with enterprise-level capabilities. ChatGPT Enterprise's offerings, like domain verification, SSO, and usage insights, directly overlap with many existing B2B services, potentially jeopardizing the survival of these startups.

In his keynote, OpenAI's CEO Sam Altman revealed another major development: the extension of GPT-4 Turbo's knowledge cutoff. Unlike GPT-4, which had information only up to 2021, GPT-4 Turbo is updated with knowledge up until April 2023, marking a significant step forward in the AI's relevance and applicability.

ChatGPT Enterprise stands out with features like enhanced security and privacy, high-speed access to GPT-4, and extended context windows for longer inputs. Its advanced data analysis capabilities, customization options, and removal of usage caps make it a superior choice to its predecessors. Its ability to process longer inputs and files, along with unlimited access to advanced data analysis tools like the previously known Code Interpreter, further solidifies its appeal, especially among businesses previously hesitant due to data security concerns.

The era of manually crafting code is giving way to AI-driven systems, trained instead of programmed, signifying a fundamental change in software development.

The mundane tasks of programming may soon fall to AI, reducing the need for deep coding expertise. Tools like GitHub's CoPilot and Replit’s Ghostwriter, which assist in coding, are early indicators of AI's expanding role in programming, suggesting a future where AI extends beyond assistance to fully managing the programming process. Imagine the common scenario where a programmer forgets the syntax for reversing a list in a particular language. Instead of a search through online forums and articles, CoPilot offers immediate assistance, keeping the programmer focused towards to goal.

Transitioning from Low-Code to AI-Driven Development​

Low-code & No code tools simplified the programming process, automating the creation of basic coding blocks and liberating developers to focus on creative aspects of their projects. But as we step into this new AI wave, the landscape changes further. The simplicity of user interfaces and the ability to generate code through straightforward commands like “Build me a website to do X” is revolutionizing the process.

AI's influence in programming is already huge. Similar to how early computer scientists transitioned from a focus on electrical engineering to more abstract concepts, future programmers may view detailed coding as obsolete. The rapid advancements in AI, are not limitd to text/code generation. In areas like image generation diffusion model like Runway ML, DALL-E 3, shows massive improvements. Just see the below tweet by Runway showcasing their latest feature.



Extending beyond programming, AI's impact on creative industries is set to be equally transformative. Jeff Katzenberg, a titan in the film industry and former chairman of Walt Disney Studios, has predicted that AI will significantly reduce the cost of producing animated films. According to a recent article from Bloomberg Katzenberg foresees a drastic 90% reduction in costs. This can include automating labor-intensive tasks such as in-betweening in traditional animation, rendering scenes, and even assisting with creative processes like character design and storyboarding.

The Cost-Effectiveness of AI in Coding​

Cost Analysis of Employing a Software Engineer:

1. Total Compensation: The average salary for a software engineer including additional benifits in tech hubs like Silicon Valley or Seattle is approximately $312,000 per year.

Daily Cost Analysis:

1. Working Days Per Year: Considering there are roughly 260 working days in a year, the daily cost of employing a software engineer is around $1,200.
2. Code Output: Assuming a generous estimate of 100 finalized, tested, reviewed, and approved lines of code per day, this daily output is the basis for comparison.

Cost Analysis of Using GPT-3 for Code Generation:

1. Token Cost: The cost of using GPT-3, at the time of the video, was about $0.02 for every 1,000 tokens.
2. Tokens Per Line of Code: On average, a line of code can be estimated to contain around 10 tokens.
3. Cost for 100 Lines of Code: Therefore, the cost to generate 100 lines of code (or 1,000 tokens) using GPT-3 would be around $0.12.

Comparative Analysis:

• Cost per Line of Code (Human vs. AI): Comparing the costs, generating 100 lines of code per day costs $1,200 when done by a human software engineer, as opposed to just $0.12 using GPT-3.
• Cost Factor: This represents a cost factor difference of about 10,000 times, with AI being substantially cheaper.

This analysis points to the economical potential of AI in the field of programming. The low cost of AI-generated code compared to the high expense of human developers suggests a future where AI could become the preferred method for code generation, especially for standard or repetitive tasks. This shift could lead to significant cost savings for companies and a reevaluation of the role of human programmers, potentially focusing their skills on more complex, creative, or oversight tasks that AI cannot yet handle.

ChatGPT's versatility extends to a variety of programming contexts, including complex interactions with web development frameworks. Consider a scenario where a developer is working with React, a popular JavaScript library for building user interfaces. Traditionally, this task would involve delving into extensive documentation and community-provided examples, especially when dealing with intricate components or state management.

With ChatGPT, this process becomes streamlined. The developer can simply describe the functionality they aim to implement in React, and ChatGPT provides relevant, ready-to-use code snippets. This could range from setting up a basic component structure to more advanced features like managing state with hooks or integrating with external APIs. By reducing the time spent on research and trial-and-error, ChatGPT enhances efficiency and accelerates project development in web development contexts.

Challenges in AI-Driven Programming​

As AI continues to reshape the programming landscape, it’s essential to recognize the limitations and challenges that come with relying solely on AI for programming tasks. These challenges underscore the need for a balanced approach that leverages AI's strengths while acknowledging its limitations.

1. Code Quality and Maintainability: AI-generated code can sometimes be verbose or inefficient, potentially leading to maintenance challenges. While AI can write functional code, ensuring that this code adheres to best practices for readability, efficiency, and maintainability remains a human-driven task.
2. Debugging and Error Handling: AI systems can generate code quickly, but they don't always excel at debugging or understanding nuanced errors in existing code. The subtleties of debugging, particularly in large, complex systems, often require a human's nuanced understanding and experience.
3. Reliance on Training Data: The effectiveness of AI in programming is largely dependent on the quality and breadth of its training data. If the training data lacks examples of certain bugs, patterns, or scenarios, the AI’s ability to handle these situations is compromised.
4. Ethical and Security Concerns: With AI taking a more prominent role in coding, ethical and security concerns arise, especially around data privacy and the potential for biases in AI-generated code. Ensuring ethical use and addressing these biases is crucial for the responsible development of AI-driven programming tools.

Balancing AI and Traditional Programming Skills​

In future software development teams maybe a hybrid model emerges. Product managers could translate requirements into directives for AI code generators. Human oversight might still be necessary for quality assurance, but the focus would shift from writing and maintaining code to verifying and fine-tuning AI-generated outputs. This change suggests a diminishing emphasis on traditional coding principles like modularity and abstraction, as AI-generated code need not adhere to human-centric maintenance standards.

In this new age, the role of engineers and computer scientists will transform significantly. They'll interact with LLM, providing training data and examples to achieve tasks, shifting the focus from intricate coding to strategically working with AI models.

The basic computation unit will shift from traditional processors to massive, pre-trained LLM models, marking a departure from predictable, static processes to dynamic, adaptive AI agents.

The focus is transitioning from creating and understanding programs to guiding AI models, redefining the roles of computer scientists and engineers and reshaping our interaction with technology.

The Ongoing Need for Human Insight in AI-Generated Code​

The future of programming is less about coding and more about directing the intelligence that will drive our technological world.

The belief that natural language processing by AI can fully replace the precision and complexity of formal mathematical notations and traditional programming is, at best, premature. The shift towards AI in programming does not eliminate the need for the rigor and precision that only formal programming and mathematical skills can provide.

Moreover, the challenge of testing AI-generated code for problems that haven't been solved before remains significant. Techniques like property-based testing require a deep understanding programming, skills that AI, in its current state, cannot replicate or replace.

In summary, while AI promises to automate many aspects of programming, the human element remains crucial, particularly in areas requiring creativity, complex problem-solving, and ethical oversight.

 

[Traveler]

I just started using chatgpt and canva to handle all my wife's cleaning company social media posts. Saves me souch time and money since I don't need a social media manager

 

[bnew]

GitHub - abi/screenshot-to-code: Drop in a screenshot and convert it to clean code (HTML/Tailwind/React/Vue)

Drop in a screenshot and convert it to clean code (HTML/Tailwind/React/Vue) - abi/screenshot-to-code

github.com