dtflare (@artificial_bug) header
dtflare (@artificial_bug) profile picture

Faster than AI. Cybersecurity | AI

Joined September 2023
Photo by dtflare (@artificial_bug)Photo by dtflare (@artificial_bug)Photo by dtflare (@artificial_bug)Photo by dtflare (@artificial_bug)Photo by dtflare (@artificial_bug)Photo by dtflare (@artificial_bug)Photo by dtflare (@artificial_bug)Photo by dtflare (@artificial_bug)Photo by dtflare (@artificial_bug)Photo by dtflare (@artificial_bug)
Logo design process for ‘Simones Bar & Lounge’ a high end and premium club situated in the downtown area of the city. This place will serve the imported liquor and great quality food. . A sharp, minimal and premium looking logo was required for this place
13
31
1
412
50,807
Introducing the 01 Developer Preview. Order or build your own today: openinterpreter.com/01 The 01 Light is a portable voice interface that controls your home computer. It can see your screen, use your apps, and learn new skills. This is only the beginning for 01— the open-source foundation for this new era of AI devices.
573
1,056
833
5,643
1,409,692
Here is Claude 3's system prompt! Let me break it down 🧵
112
530
219
3,305
985,516
X + Midjourney talks are underway with the goal of a partnership. Big if true. Let's imagine for a second, shall we? > Collaborative Midjourney generation on X would be wild — like Discord parties on steroids. No blank screen problem because prompts can be remixed, content can evolve, with full provenance. You can see memes being birthed in real time — like photoshop tennis on steroids. It would inject a stream of visual umami through the X network. > xAI tie-in to tackle multimodal AI creation and assistance. X brings some LLM and vision chops to the party to go up against Runway, OpenAI, Pika, Google, etc. Give @DavidSacks and real content creators the controlled creation experience we so badly desire —versus playing slot machine AI all day re-rolling prompts and picking a few to post on social media. Memetic warfare demands real weapons, so they must be built. Which also means… > Deep partnership on identifying generative content on X. Deepfakes and bots have to be top of mind for @elonmusk given the upcoming elections. Midjourney is one of the best image generators out there and let you generate a variety of political figures and celebrities, creating very convincing imagery often indistinguishable from reality. > What if anyone on X could make an AI avatar profile picture? ICYMI, X recently updated ToS to collect your biometric data so this should be easy. Expand to include custom emotes and stickers with more data. So many other ways to imagine Midjourney integration into the X creation experience, which is pretty parse at the moment and this could be nice boost. > Speaking of boost, compute and data. GPUs aren’t cheap and everyone’s gotta pay the Jensen tax. Could there be a collaboration on shared compute and perhaps even jointly training models. Real-time and historic data from X could do some pretty interesting things, especially now with a growing library of longer form media. That concludes my speculation on what a Midjourney and X partnership might look like. I look forward to seeing how this evolves because it could be dope, if true. Enjoyed this? You might also enjoy my feed on all things AI & sPaTiAL cOmPuTiNg: @bilawalsidhu
BREAKING: 𝕏 is in talks with Midjourney for a potential partnership.
22
18
7
92
20,923
dtflare retweeted
Nomic Embed Training a Reproducible Long Context Text Embedder paper page: huggingface.co/papers/2402.0… This technical report describes the training of nomic-embed-text-v1, the first fully reproducible, open-source, open-weights, open-data, 8192 context length English text embedding model that outperforms both OpenAI Ada-002 and OpenAI text-embedding-3-small on short and long-context tasks. We release the training code and model weights under an Apache 2 license. In contrast with other open-source models, we release a training data loader with 235 million curated text pairs that allows for the full replication of nomic-embed-text-v1.
6
38
4
181
34,369
Today I have a huge announcement. The dataset used to create Open Hermes 2.5 and Nous-Hermes 2 is now PUBLIC! Available Here: huggingface.co/datasets/tekn… This dataset was the culmination of all my work on curating, filtering, and generating datasets, with over 1M Examples from datasets from across the open source ecosystem and some that I generated as well. Super excited to be able to finally share this with you all and even more excited to see what you all make from it's release! Every data source (except one that I can no longer find) is attributed in the data card, but I will post it here in this thread as well - If they have a twitter account they'll be tagged so be sure to follow them all!
100
284
60
1,892
237,190
Save up to 8x on embeddings. We can’t wait to see what you build on Together AI. api.together.ai
14
25
16
274
90,649
🥱 LazyMergekit I made a Colab notebook to automate model merging using mergekit. It's very convenient, only requires a CPU, and allows you to create MoE like Beyonder. 💻 Colab: colab.research.google.com/dr…
4
58
305
25,301
Mark Zuckerberg has the right strategy. It's just that his device has 2K screens. Apple's has 5K screens. Which one do you want to answer emails on?
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
258
72
16
835
463,850
WhiteRabbitNeo "WRN-Chapter-1" dataset for cybersecurity is now expanded to contain 7750 super high-quality samples. This will help most open source AIs to offer advanced cybersecurity capabilities. Enjoy! huggingface.co/datasets/whit…
3
16
79
7,286
dtflare retweeted
Embedding features learned with sparse autoencoders can make semantic edits to text ✨ (+ a reading/highlighting demo) I've built an interface to explore and visualize GPT-4 labelled features learned from a text embedding model's latent space. Here's a little video, more in 👇
14
36
7
357
52,721
dtflare retweeted
I wrote about our investment rationale into Sakana AI. We are excited to partner with @hardmaru and @YesThisIsLion on their ambitious vision to push the bounds of global AI. Congrats to @SakanaAILabs on the seed round! 500.co/theglobalvc/sakana-ai… @500GlobalVC
Excited to announce our seed round! We raised $30M to develop nature-inspired AI in Japan. sakana.ai/seed-round
2
5
1
18
11,863
South Korea is making their bet on what the future economy will look like, and investing almost half a trillion USD in 13 new chip plants. Construction is starting now, with a plan to continue over the next twenty-five years, aiming to establish the world's largest chip hub.
16
28
11
230
29,694
I just shipped the v0.6 release of llamafile. It supports multiple GPUs, AMD GPUs, the OpenAI API server is more reliable, and the help docs are getting better. github.com/Mozilla-Ocho/llam…
8
57
2
390
30,236
New Anthropic Paper: Sleeper Agents. We trained LLMs to act secretly malicious. We found that, despite our best efforts at alignment training, deception still slipped through. arxiv.org/abs/2401.05566
105
527
317
2,906
1,844,419
I touched on the idea of sleeper agent LLMs at the end of my recent video, as a likely major security challenge for LLMs (perhaps more devious than prompt injection). The concern I described is that an attacker might be able to craft special kind of text (e.g. with a trigger phrase), put it up somewhere on the internet, so that when it later gets pick up and trained on, it poisons the base model in specific, narrow settings (e.g. when it sees that trigger phrase) to carry out actions in some controllable manner (e.g. jailbreak, or data exfiltration). Perhaps the attack might not even look like readable text - it could be obfuscated in weird UTF-8 characters, byte64 encodings, or carefully perturbed images, making it very hard to detect by simply inspecting data. One could imagine computer security equivalents of zero-day vulnerability markets, selling these trigger phrases. To my knowledge the above attack hasn't been convincingly demonstrated yet. This paper studies a similar (slightly weaker?) setting, showing that given some (potentially poisoned) model, you can't "make it safe" just by applying the current/standard safety finetuning. The model doesn't learn to become safe across the board and can continue to misbehave in narrow ways that potentially only the attacker knows how to exploit. Here, the attack hides in the model weights instead of hiding in some data, so the more direct attack here looks like someone releasing a (secretly poisoned) open weights model, which others pick up, finetune and deploy, only to become secretly vulnerable. Well-worth studying directions in LLM security and expecting a lot more to follow.
New Anthropic Paper: Sleeper Agents. We trained LLMs to act secretly malicious. We found that, despite our best efforts at alignment training, deception still slipped through. arxiv.org/abs/2401.05566
164
665
111
4,810
908,286
Introducing r1. Watch the keynote. Order now: rabbit.tech #CES2024
1,261
2,330
3,146
13,082
11,125,438
Research Idea 1 2024: the little prince has been translated into more languages than any other book except the bible (505 languages). and the book has entered the public domain -- I am surprised no-one has structured it into a instruction style translation dataset?
5
11
1
134
29,397
dtflare retweeted
Following up from the great work from community that enabled bitsandbytes 4-bit serialization, I pushed Mixtral-Instruct-bnb-4bit on @huggingface for anyone that wants to easily load the model huggingface.co/ybelkada/Mixt…
7
27
1
228
49,671
To use SDXL, we need a web interface. Some popular choices are: 1. Automatic1111 - Web UI 2. Comfyanonymous - ComfyUI 3. lllyasviel - Fooocus Fooocus is the simpler one, plus it has all the features we need at the moment. So let's install Fooocus.
3
12
1
104
75,818