3511
Claude 2.1
Claude 2.1 brings a 200K token context window, a 2x reduction in hallucination rates, beta tool use, and system prompts.
New
The latest blog posts to take off on Hacker News. 6,334 posts.
3512
Margaret Mead, technocracy, and the origins of AI's ideological divide
The anthropologist helped popularize both techno-optimism and the concept of existential risk
3513
3514
Getting Started with Mistral-7b-Instruct-v0.1
Fast and Portable Llama2 Inference on the Heterogeneous Edge.
3515
3516
OpenAI: Facts from a Weekend
Approximately four GPTs and seven years ago, OpenAI’s founders brought forth on this corporate landscape a new entity, conceived in liberty, and dedicated to the proposition that all men might live…
3517
3518
Microsoft is eating the world
If OpenAI was a public company, it's stock would have been down about 38% today in a shocking turn of events.
3519
OpenAI's misalignment and Microsoft's gain
The end of a dramatic weekend in tech is that OpenAI has split and Microsoft is partnered with one and has hired the other; this is the ultimate failure case of what should have been a for-profit c…
3520
Practical Tips for Finetuning LLMs Using LoRA (Low-Rank Adaptation)
Things I Learned From Hundreds of Experiments
3521
Bootstrapping self awareness in GPT-4: Towards recursive self inquiry
Or, Look Ma – I Made GPT-4 Self Aware!
3522
The architecture of today's LLM applications
Here’s everything you need to know to build your first LLM app and problem spaces you can start exploring today.
3523
A Timeline of the OpenAI Board
Yesterday, Sam Altman and Greg Brockman were fired from the Board of Directors of OpenAI.
3524
I disagree with Geoff Hinton regarding "glorified autocomplete"
“People say, It’s just glorified autocomplete . . . Now, let’s analyze that. Suppose you want to be really good at predicting the next word. If you want to be really good, you have to understand what’s being said. That’s the only way. So by training something to be really good at predicting the next word, you’re…
3525
A PCIe Coral TPU Finally Works on Raspberry Pi 5
Coral.ai TPUs are AI accelerators used for tasks like machine vision and audio processing. Raspberry Pis are often integrated into small robotics and IoT products—or used to analyze live video feeds with Frigate. Until today, nobody I know of has been able to get a PCI Express Coral TPU working on the Raspberry Pi…
3526
Emu Video and Emu Edit, our latest generative AI research milestones
Today, we’re announcing new research into controlled image editing based solely on text instructions and a method for text-to-video generation based on diffusion models.
3527
A failed AI girlfriend product, and my lessons
Lessons from Dolores, my failed AI girlfriend product: built on the Stanford AI Town paper with GPT and ElevenLabs voices, why it earned $50 and I gav
3528
Google's advanced music generation model and two new AI experiments
Announcing our most advanced music generation model and two new AI experiments, designed to open a new playground for creativity
3529
Exploring GPTs: ChatGPT in a trench coat?
The biggest announcement from last week’s OpenAI DevDay (and there were a LOT of announcements) was GPTs. Users of ChatGPT Plus can now create their own, custom GPT chat bots …
3530
Writing a GPT-4 script to check Wikipedia for the first unused acronym
Curious what the first ‘unused’ alphabetic acronym is, I have GPT-4 write a script to check English Wikipedia. After three bugs, the first unused one turns out as of 2023-09-29 to be the three-letter acronym ‘CQK’, with another 2.6k TLA unused, and 393k four-letter acronyms unused. Exploratory analysis suggests…
3531
Is the reversal curse in LLMs real?
A recent paper The Reversal Curse points out an apparent failure in large large language models like GPT-4. From the abstract: We expose a surprising failure of generalization in auto-regressive large language models (LLMs). If a model is
3532
GraphCast: AI model for weather forecasting
Our state-of-the-art model delivers 10-day weather predictions at unprecedented accuracy in under one minute
3533
Infinite Context LLMs: Going Beyond RAG with Extended Minds
In this blog we discuss how the transformer architecture naturally extends over external memories, and share empirical results which leverage this capability. These methods are innate (don't require fine tuning) and outperform popular retrieval augmented generation methods.
3534
Hacking Google Bard – From Prompt Injection to Data Exfiltration
Recently Google Bard got some powerful updates, including Extensions. Extensions allow Bard to access YouTube, search for flights and hotels, and also to access …
3535
3536
Fast and Portable Llama2 Inference on the Heterogeneous Edge
Fast and Portable Llama2 Inference on the Heterogeneous Edge.
3537
GPU Survival Toolkit for the AI age
Why CPU Knowledge Is No Longer Enough In today's AI age, the majority of developers train in the CPU way. This knowledge has been part of our academics as well, so it's obvious to think and problem-solve in a CPU-oriented way. However, the problem with CPUs is that they rely
3538
Google Cloud TPU Multislice Training
We used Multislice Training to run the world’s largest LLM distributed training job on a compute cluster of 50,944 Cloud TPU v5e chips.
3539
Don't build AI products the way everyone else is doing it
More and more, people are building products as thin wrappers over other models, especially LLMs. This can cause major issues, and there is a better way.
3540
We're sorry we created the Torment Nexus
( This is the text of a talk I delivered at the Next Frontiers Applied Fiction Day in Stuttgart on Friday November 10th, 2023. Note: early draft, contains some typos, I'll fix them next week when I get home. )