ai Twitter List: August 19, 2024 AM Edition
Length: • 11 mins
Annotated by Jimmy Lv
gm. what will you ship?
🧙🏻♂️ Dify v0.7.1 is here with OpenAI Structured Outputs support and customized app icon, and more:
🔢 OpenAI Structured Outputs (gpt-4o-mini & gpt-4o-2024-08-06 only)
When building applications, select OpenAI models, then enable and customize JSON Schema in the LLM node of the Workflow. This allows the LLM to output data in a predefined format, making it easier for downstream nodes to parse, process, and transform.
🖼️ Customized Application Icon
Easily upload your image to customize your app icon when you start to build, making it more personalized and interesting.
📦 Difysandbox has been updated to v0.2.6, relaxing restrictions to allow more freedom in installing Python packages.
⚙️ New Models and Tools
- chatgpt-4o-latest: a chat-optimized model
- Jina Tokenizer: converts text into tokens or chunk long documents into smaller segments easier for LLM processes, and the API is free!
- GitLab: helps you directly query and retrieve file contents and code commit records from GitLab within Dify
🐞 Plus many bug fixes in this release thanks to our community.
See full changelog here: https://t.co/btQioLyL9n
I see lots of people making very strong and demanding claims about the nature of "reasoning" in order to exclude LLMs. But in setting the goalposts so high, this approach risks the absurd conclusion that most humans rarely if ever engage in reasoning themselves.
When I ended up on the street in 2018, I lost everything. I couldn't go home. My wife slapped a restraining order on me to protect her and the kids from my behaviors, and I was badly addicted to heroin and later fentanyl. I would daydream of getting my own SRO (apartment) from the city so I could have a place to use drugs alone.
Thinking back, if that had happened, I'd be dead now because I would've been isolated (which is what you want in addiction) and that would have been the worst possible outcome. I'm not against housing people. I support more housing. But ever since fentanyl arrived on the streets, everything changed. Now, 70% of all OD deaths happen at a fixed address in San Francisco. Bodies are constantly pulled out of SROs and homeless hotels. Yet, we continue to downplay this truth and double down on existing policy that was put in place before fentanyl arrived. The drug of choice has changed from heroin (which you could use and potentially survive on for years) to illicit fentanyl that's 10-100x stronger. Yet, we continue on with the same approach that simply no longer works for too many people stuck in addiction.
Times change. Drugs change. The drug dealers change. But our policies have not changed. Now, we see the results, and it has sparked a backlash because we are horrified at the amount of death we are witnessing in our city and our country because of fentanyl. There are 108,000 people dying every year from drugs. It's the leading cause of death for people under 55. Yet, we treat it like suicide or a terminal disease. Addiction is treatable! As with any disease, you want to treat it to make it better, not maintain your disease where it's at. We must stop the flow of fentanyl AND invest in treatment. It's never one or the other. It's both. Supply and demand. We will keep failing if we keep the pendulum swinging from one extreme to the other. In the case of addiction and homelessness, we must land in the center. @KamalaHarris @realDonaldTrump @RobertKennedyJr
@mccaffreyr3 @ONDCP @DEAHQ
Creative news you *need* to know:
🎬 @DescriptApp released collaborative recording “Rooms,” a Zoom Integration, and an AI-powered automatic multi-camera feature.
⚡️ @LottieLabHQ released “Interactivity,” which allows designers to create and adjust animation states, add interactive elements, and more in a single editor.
📸 @HalideCamera released “Process Zero,” a new mode that removes all AI and standard image processing from your photos.
📊 @NotionHQ released “Charts,” allowing you to turn any database into a chart with just one click.
❏ @Webflow now supports drop shadows as a CSS effect in the style panel.
🎉 @Midjourney released a new, updated web editor.
📝 @NotionHQ is launching Forms later this summer.
📄 @Canva released five updates to Docs, including page outlines, highlight blocks, and page breaks.
🤳 @Google released a new “Add Me” feature to the Google Pixel, which makes taking group photos easier.
🎙️ @ElevenLabsio released “ElevenStudios,” a fully managed AI dubbing service currently working with Colin and Samir, Ali Abdaal, and others.
⏩ @RunwayML Gen-3 Alpha Turbo Image-to-Video is now available and can generate videos 7x faster for half the price of Gen-3 Alpha.
📦 @SplineTool released native 3D, real-time API & webhooks, and more for Android.
🆕 @Elgato now allows you to add custom Adobe Photoshop shortcuts to your Stream Deck.
🤝 @CircleApp released “Circle 3.0,” a new update that includes new gamification, layouts, and more.
😳 @Patreon will now take an additional 30% of creator earnings due to Apple’s in-app purchase commission.
🖱️ @Rive_App now allows users to use the "Click" option for interactions, allowing them to cancel actions by not completing the full "Click" action.
🎧 @Bezi_3D now allows you to add audio to your projects.
♾ @HeyGen_Official released a new unlimited Creator and Team plan allowing you to create unlimited videos.
🔟 @LumaLabsAI's “Dream Machine” reached 10 million users in just seven weeks.
👀 @TryRamp released a “Summer 2024 Spend Report” with insights on the top AI tools companies pay for.
🖼️ @X released Grok 2.0, which allows you to generate images using the popular Flux.1 AI model.
🆕 @Framer will release a new app icon soon.
🎁 @LumaLabsAI’s “Dream Machine” v1.5 is launching next week.
👏 @Rive_App removed the file limit on personal files.
NEW: Johns Hopkins University will go back to requiring standardized test scores for admission in 2026:
“With input from faculty colleagues, the review concluded that test scores, when considered in context as part of a holistic approach to admissions, serve as an important predictive metric to assess the likelihood of a student's academic success at Johns Hopkins.”
DPIs - it takes 10-12 years for a company to go from being founded to IPO. Don't expect much DPI in the first 5 years of an early-stage fund's life. Also, liquidity in VC comes in waves. 2021 was one of these waves. So, no surprise later vintages are lagging on DPI. /5
Looks like my first 10 mil impression tweet coming soon.
✅ achievement unlocked
“Neo, we learned it from you”
Make kindness your default setting
Check out the latest podcast with Groq Fellow, @satnam6502. He discusses the value of specification, the current state of academia, gaming the metrics, functional programming in hardware, bullying, and more.
hubs.la/Q02LwLv70
AI TPOT is a bubble. We might think everyone knows the latest tools or techniques just because they fill out timelines, but the reality is that 99% of the world is yet to be AI-pilled.
Friend of mine, former SWE at FAANG.
7 spots have been purchased! 😭❤️👏
It all started with Reid setting the example. Then an anonymous leader in this space came in for 5. Then @webflow covered a spot. This is why this community is so special. We come together for others!
I have 2 left if anyone is interested.
I’d like to GIVE most of these spots to designers that can’t afford to come but need the connections and inspiration.
Any sponsors want to come in on $10-20K and help change the careers of some amazing designers? Maybe end up hiring one?
Which are you choosing? 👀
From idea to production without writing any code.
Well done, @akhi20187 !
Today, I have launched smoketowealth.replit.app that tells how much an individual can earn, had they quit smoking and start investing. Built without writing code. Python, ReactJS
@claudeai101 -Code
@Replit -Host
@Investingcom - Data
@rileybrown_ai , @100xengineers - Inspiration
The best entrepreneurs are those who are willing to learn, adapt, and continuously innovate
The wait is finally over!!! 😁
We just dropped an in-depth tutorial on how to build your own robot!
Teach it new skills by showing it a few moves with just a laptop.
Then watch your homemade robot act autonomously 🤯
1/🧵👇
What a sight! Thanks @NYSE for celebrating our place on the @Forbes Cloud 100 list ☁
"The competitor to be feared is one who never bothers about you at all, but goes on making his own business better all the time."
— Henry Ford
🚀 Introducing Hugging Face Similar: a Chrome extension to find relevant datasets!
✨ Adds a "Similar Datasets" section to @huggingface dataset pages
🔍 Recommendations based on dataset READMEs
🏗️ Powered by @trychroma and @SnowflakeDB
Get it now!
chromewebstore.google.com/detail/hugging...
DSPy got @langfuse integration
A @github Copilot-guided experience for database migrations
Temu sellers are unstoppable at satisfying American revealed preferences.
The obvious thing is for Apple and Google and FB to autorun a fine tune on all your photos on their service and allow you to prompt any photo of yourself that you want...
But the safety and PR and political fears keep them from doing it.
Wonder how long the truce will hold
I trained a LoRa with FLUX on a few photos of me and brought it to life with Runway.
This image AND video is 100% AI-generated.
Here's a couple more wild examples and a full tutorial for how you can do it yourself (it's addicting):
Me: Trying to explain why I didn't hit the growth goal in YC group session
Welcome SmolLM Instruct v0.2 - 135M, 360M and 1.7B parameter instruction tuned Small LMs, Apache 2.0 licensed. Closing the gap to bring intelligence closer to thought (<500 ms per generation)! 🔥
The models are optimised to run on-device with WebGPU support (from MLC and ONNX Runtime) and llama.cpp.
Run them on your Mac, browser, GPU, CPU - it works blazingly fast.
We provide already converted/ quantised - GGUFs, MLC and ONNX checkpoints 🐐
What's new?
We train SmolLM base models on a new synthetic dataset of 2,000 simple everyday conversations we generated by llama3.1-70B -> everyday-conversations-llama3.1-2k
and existing datasets like Magpie-Pro-300K-Filtered by @argilla_io, self-oss-instruct-sc2-exec-filter-50k, and a small subset of OpenHermes-2.5 from @NousResearch
Bonus: We release the fine-tuning scripts we used to train these checkpoints, so that you can fine-tune them for your own use-cases too. ⚡
Enjoy! and looking forward to what you build with these 🤗
This app is free
How to Build a Simplified OpenAI Memory using Dify Conversation Variables
Conversation Variables are short-term memory units employed by Dify to provide temporary storage in multi-turn conversations within chatflows. This blog will walk you through how to use them to emulate OpenAI's Memory Features and there's also a ready-to-use template to get you started. Check it out!
San Francisco dominance in venture capital is growing 📈
“AI stuff feels like it’s going to be dominated by the Big Tech companies in the San Francisco Bay Area”
Cc: @garrytan @erikbryn @Noahpinion @alexandr_wang
Interesting new paper on automating the process of building powerful AI agents.
Presents Meta Agent Search a meta agent that iteratively programs and tests new agents based on a growing archive of previous discoveries.
Claims that with their approach it is possible to learn any possible agentic system including prompts, tool use, control flows, and more.
They achieve this by focusing on three main components referred to as search space (define agents), search algorithm (explore search space), and the evaluation function (evaluate candidate agents).
"We consistently observe the surprising result that agents invented by Meta Agent Search maintain superior performance even when transferred across domains and models, demonstrating their robustness and generality."
Just imagine that mkbhd can think 💭 this, and seconds later a bot in Shenzhen send an order “10 large by Monday?” “Confirmed”, the image goes up on Temu, and seconds later orders are pouring in, factories are running, plastic is flowing, packages are shipped and a week later a parent in Minnetonka opens a corrugated cardboard box scratches his head and is bewildered by the tastes of his offspring is it a joke an amorous toy
My friend has an incredible phone case 😭
Enhancing Greedy Decoding with LLMs using Diverse Exemplars
Uses a hybrid self-ensembling approach (based on diverse exemplars) to improve the overall performance of LLMS.
Specifically, it uses diverse exemplars to generate multiple candidate responses and then aggregates them using an LLM to generate a final response.
This approach achieves better accuracy compared to greedy decoding and lower cost compared to self-consistency approaches.
I was sent stats for the YC board meeting tomorrow. The second number is the fraction of companies with no revenue when YC funded them. High is good because it means we're investing early. If this doesn't convince you that you don't have to wait to apply, I don't know what will.
Launching nukes into the global consciousness
Watching ripples from the point impact
Butterflies hah
Basically, much of America has decided not to investigate murders when the victim is Black. I suspect this is sort of a devil's bargain between those who want less policing and those who just don't care when Black people die.
A weird detail about this is that it's a race-specific trend.
I really like Kamala Harris' housing plan. It takes us a little bit in the direction of Singapore's system, which I view as the best system out there.
This could sink the Democratic convention. The Democrats need the moderates to win. They won't get them with this. Just sayin...