[412-code-talk] Fwd: GPT-4 is exciting and scary ⁉️, Adept raises $350M 💰, PyTorch 2.0 release 2️⃣

Thu Mar 16 10:03:38 EDT 2023

---------- Forwarded message ---------
From: TLDR AI <dan at tldrnewsletter.com>
Date: Thu, Mar 16, 2023 at 9:06 AM
Subject: GPT-4 is exciting and scary ⁉️, Adept raises $350M 💰, PyTorch 2.0
release 2️⃣
To: <mark at rauterkus.com>

GPT-4 also scored high on a variety of tests, including the Uniform Bar
Exam and the Biology Olympiad.

Sign Up
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Ftldr.tech%2Fai%3Futm_source=tldr/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/QJUwyfIUGW3YWgJOZdMU1ReVFvUPYjyAbMZzRDz-JlM=292>
|Jobs
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Ftldr.tech%2Fjobs/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/sin6k0GV7W2-Rt2bs6-qnqqnxakQKAklPZwoPUqh39M=292>
|Advertise
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Fshare.hsforms.com%2F1OxvmrkcFS4qsxKpNXCi76wee466%3Futm_source=tldrai%26utm_medium=newsletter/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/JE52m3Ndc9-hzwiEccFFJjO-8zRvGJ3EzodI4rIChgY=292>
|View Online
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Factions.tldrnewsletter.com%2Fweb-version%3Fep=1%26lc=078d99d6-b44d-11ed-ba38-55928061a93d%26p=40b89f60-c3d6-11ed-b8a4-a13e1db6e66a%26pt=campaign%26t=1678971989%26s=72474782a9dbf58d4dfc0415d6f3f177a8e423c891b11df72c8a59c936ce0d34/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/Q7zPupB_uWB3S_9H3EG8Pjkek7Ov8ZMAZk-fomgjYwc=292>
TLDR
*Daily Update 2023-03-16*
🚀
*Headlines & Launches*
*GPT-4 is exciting and scary (2 minute read)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Farchive.ph%2FUDPRi%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/C2yoMDj99wB5OS1vXIaVXUm4i8SdkbcuToafEk8ccqE=292>

The article discusses the author's experience testing GPT-4 by asking it a
variety of questions — to help with a tax problem, to come up with a new
word, to plan a birthday party. The author found that GPT-4 was capable of
a wider range of tasks, more fluid in its responses, and seemed to have
more guardrails in place than its predecessor. GPT-4 also scored high on a
variety of tests, including the Uniform Bar Exam and the Biology Olympiad.
While the article acknowledges the positive potential of GPT-4, it also
discusses the negative side — the thought of AI being smarter than humans,
ethical implications of such technology, and the potential for misuse.
Also, another correction from yesterday’s newsletter: GPT-4 can take both
*images* and text as inputs, but not video. Sorry for the brain fart! Maybe
I should have had GPT-4 proofread it :)
*Adept AI raises $350M in Series B funding (2 minute read)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Fwww.adept.ai%2Fblog%2Fseries-b%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/xm5ijFxrxtaGtAbWiHp7vxVF6nZnUmZ-Do9QRID4pkQ=292>

Adept has raised $350M in new funding as part of their Series B led by
General Catalyst and Spark Capital. Adept is a research and product AI lab
that automates software processes for developers. It's similar to robotic
process automation (RPA), or software robots that use a combination of
automation, computer vision, and machine learning to automate repetitive
tasks like filing forms and responding to emails. Adept is creating general
intelligence by allowing humans and computers to collaborate creatively.
*Google Cloud brings generative AI to developers, businesses, and
governments (7 minute read)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Fcloud.google.com%2Fblog%2Fproducts%2Fai-machine-learning%2Fgenerative-ai-for-businesses-and-governments%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/M5mxW-Lf9pzffLaeeNoVc_9HF6_S0ycEHk5fmROyf5Q=292>

Google Cloud has launched Generative AI support in Vertex AI and Generative
AI App Builder, which enables businesses and governments to access and
customize generative AI technology. These tools allow developers to create
generative AI applications quickly and easily using out-of-the-box
templates and API access to Google's foundation models, with full control
over data usage. This move reflects Google Cloud's commitment to empowering
teams and transforming industries through AI technology.
🧠
*Research & Innovation*
*Reinforcement learning from Privacy Feedback (6 minute read)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Fgretel.ai%2Fblog%2Fteaching-large-language-models-to-zip-their-lips%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/iIZ4Tt_-hoE6as01Zy1QT17ZOGQs8lfdHr0NKM7tt_w=292>

Language models often leak sensitive data. This is undesirable for many use
cases, but hard to avoid due to the next token prediction training
objective. By replacing the human feedback objective in RLHF with a privacy
preserving reward, we are able to tune language models that leak much less
often than their untrained counterparts.
*UPRISE: Universal Prompt Retrieval for Improving Zero-Shot Evaluation in
Large Language Models (18 minute read)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Farxiv.org%2Fabs%2F2303.08518%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/PRJDDFJe_KsaSj7nnJ-WutX_h-7KmqgihRLqWuR-9m8=292>

This paper proposes a solution to the challenge of fine-tuning and
task-specific prompt engineering that can limit the generalization
abilities of Large Language Models (LLMs). The proposed method, UPRISE,
uses a lightweight and versatile retriever to automatically retrieve
prompts for a given zero-shot task input, and demonstrates universality in
a cross-task and cross-model scenario. The method is tested on a diverse
set of tasks, using a small frozen LLM for tuning and testing on much
larger LLMs, and is shown to mitigate the hallucination problem in
experiments with ChatGPT, suggesting potential to improve even the
strongest LLMs.
*DeepMIM: Deep Supervision for Masked Image Modeling (7 minute read)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Farxiv.org%2Fabs%2F2303.08817%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/cFQHSnbHIKriHK3uMd1JobONTJnTj7ukrUs6zODfqFc=292>

This paper introduces DeepMIM, a deep supervision technique that
significantly boosts the representation capability of each layer in Vision
Transformers for masked image modeling. The results show that DeepMIM
accelerates model convergence, expands attention diversities, and achieves
state-of-the-art performance on various downstream tasks.
🧑‍💻
*Engineering & Resources*
*Pip install pytorch 2.0 (5 minute read)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Fpytorch.org%2Fblog%2Fpytorch-2.0-release%2F%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/TSgQkHkVZNb5KSdXfFYqAcy-zomGmhgIs6ooR3vh32s=292>

Announced in December, this new version of pytorch is now stable and
available for easy installation. With a slew of new features, speed
improvements, and interfaces - it’s an exciting time to be using the most
popular deep learning framework out there.
*Edit-A-Video: A Text-to-Video Approach for Video Editing (Github Repo)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Fedit-a-video.github.io%2F%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/gC6jkyjEKMMG-DghNuYZbxJqq5PFLUuyY8Ao_VAiOEk=292>

This paper presents Edit-A-Video, a framework for video editing based on
text-to-video (TTV) models adapted from diffusion-based text-to-image (TTI)
models. The proposed framework inflates the 2D model into a 3D model, and
inverts the source video into the noise to edit with a target text prompt
and attention map injection, while also mitigating the background
inconsistency problem through a novel mask blending method. Experimental
results show the superiority of the proposed method in terms of background
consistency, text alignment, and video editing quality compared to
baselines.
🎁
*Miscellaneous*
*Google partners with Gretel to add synthetic data into Vertex AI (2 minute
read)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Fwww.businesswire.com%2Fnews%2Fhome%2F20230314005528%2Fen%2FGretel-Partners-With-Google-Cloud-to-Harness-the-Power-of-Synthetic-Data-and-Accelerate-Adoption-of-Safer-Generative-AI-in-the-Enterprise%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/-FZ8LCgf-3J3MDrJgmzLnLEGJmzXbQku7DHMYK4udF0=292>

Gretel provides a synthetic data platform that makes it easy to generate
anonymized, safe-to-share, and privacy-first synthetic data for enterprises
that are blocked by data sharing limitations or lack of data. Developers
that use Vertex AI for ML ops will now be able to use synthetic data to
meet privacy and data augmentation requirements, resulting in accelerated
ML R&D and operations.
*How Siri, Alexa, and Google Assistant Lost The AI Race (5 minute read)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Farchive.ph%2FnWMq2%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/Rr8CzEu_EHfLXrn9m8su7SR_pBRiwS_6_LTMs3ZzlQM=292>

Virtual assistants like Siri, Alexa, and Google Assistant had more than a
decade head start to become indispensable. But they were hampered by clunky
design and miscalculations, leaving room for chatbots like ChatGPT to rise.
*Here’s What It Would Take To Slow or Stop AI (9 minute read)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Fwww.jonstokes.com%2Fp%2Fheres-what-it-would-take-to-slow%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/3EwqAWMxc0VVPDLxGSsOTGmW6rAqt37iDn3T-ETLzN8=292>

In this thought experiment, the author breaks down each phase of creating
AI — training, fine-tuning, and inference — and what it would take to slow
down the development of AI at each of those phases. Ideas to slow down AI
proposed by the author include human intervention, GPU costs, policy
changes, legal restrictions, and even invasions.
⚡
*Quick Links*
*LinkedIn Is Adding AI Tools (1 minute read)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Fwww.theverge.com%2F2023%2F3%2F15%2F23640947%2Flinkedin-ai-profile-job-description-tools%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/hYEKRXUIWNMh97SWpNRI8PxO-Bj19lNk-CI5b9iOZ9c=292>

LinkedIn now has AI tools to assist with profile copy and job descriptions.
*LLaMA-rs (GitHub Repo)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Fgithub.com%2Fsetzer22%2Fllama-rs%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/YmqNFlDF8RmkX_u3WKM5NAdZvqBltIWMX7u3auRyl2o=292>

LLaMA-rs is a Rust port of LLaMA.
*GPT-4-For-Code (GitHub Repo)*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Fgithub.com%2Fanysphere%2Fgpt-4-for-code%3Futm_source=tldrai/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/Mt5nd-RR7li3b3QnSCW-U-Hei4PAG6i7m53GAyy0_OY=292>

This GitHub repo contains a few impressive examples of GPT-4 coding.
If you are in a hiring position, you may want to *hire AI talent through
our free job board*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Ftldr.tech%2Femployer%2Fsign-up/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/YfX0NwzI_11zfpBzQCxk0EcN7EjtdSPpMXXPpSQQ800=292>
.

If your company is interested in reaching an audience of AI
decision-makers, researchers, and engineers, you may want to *advertise
with us*
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Fshare.hsforms.com%2F1OxvmrkcFS4qsxKpNXCi76wee466%3Futm_source=tldrai%26utm_medium=newsletter/2/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/6rDO5cyN22_F4YLNVBDqyaLmp6MgrpT_GH9q4N4Im8Y=292>
.

If you have any comments or feedback, just respond to this email!

Thanks for reading,
Andrew Tan
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Ftwitter.com%2Fandrewztan/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/-1dM_BxXQOHZ97g_qZYp2YOlNDzRhWvbpnf05VmAl5I=292>
& Andrew Carr
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Ftwitter.com%2Fandrew_n_carr/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/8lIfv5wZoPQHI3wAFlv5xkWHKI_9FdT6kFtqjrkDqjY=292>

If you don't want to receive future editions of TLDR AI, please click here
to unsubscribe
<https://tracking.tldrnewsletter.com/CL0/https:%2F%2Factions.tldrnewsletter.com%2Funsubscribe%3Fep=1%26l=eedf6b14-3de3-11ed-9a32-0241b9615763%26lc=078d99d6-b44d-11ed-ba38-55928061a93d%26p=40b89f60-c3d6-11ed-b8a4-a13e1db6e66a%26pt=campaign%26pv=4%26spa=1678971618%26t=1678971989%26s=d11ad5f0ff81e6e162eb555808484bf236366fbb14ab5ccfcee181b7fbca9954/1/01000186ea876fee-2391c1d8-6627-4919-8976-46f852874ab5-000000/0KyXonfXaRBbadjFEiUoREcZ7SkgcD6NMH0OWVtnE2A=292>.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://pairlist10.pair.net/pipermail/412-pals-talk/attachments/20230316/1a14f921/attachment.htm>