I bookmarked this comment from archerx on Hacker News in this Gemma3 post.
I have tried a lot of local models. I have 656GB of them on my computer so I have experience with a diverse array of LLMs. Gemma has been nothing to write home about and has been disappointing every single time I have used it.
Models that are worth writing home about are;
Here's a one-liner to download the MNIST dataset as PNG images, white over black.
python -c "from tensorflow.keras.datasets import mnist; import os, matplotlib.pyplot as plt; (x_train, y_train), _ = mnist.load_data(); os.makedirs('mnist_images', exist_ok=True); [plt.imsave(f'mnist_images/{i}_{y}.png', img, cmap='gray') for i, (img, y) in enumerate(zip(x_train, y_train))]"
Run this in the command-line interface with a Python environment with the tensorflow and matplotlib packages.
I have two Synology DS923+ NAS devices, one running at home and the other remotely. I backup specific folders of each NAS to its remote location with Hyper Backup and Tailscale, running DSM 7.2.21. Tailscale puts both machines in the same network, removing the need to expose my networks publicly and letting Hyper Backup find the remote machine by its Tailscale IP address. This workflow works great, but yesterday I noticed Hyper Backup couldn't find the remote machine, marking it as Off-line.
Before you test my solution below, try restarting both machines from the DSM interface.
If that doesn't work, here's how I fixed the issue and restored my backups.
I uninstalled and reinstalled Tailscale on both Synology DSMs and restarting the machines. Note that this changes the Tailscale IP address of the machine and requires reauthentication, which isn't straightforward for a remote machine because you'll lose the ability to connect via SSH (the machine isn't in the Tailscale network anymore). I had to run scheduled tasks with logging on to (a) enable Tailscale, (b) update it, (c) start it, and then (d) see the authentication URL for me to log into the browser.
These are the commands I triggered manually, and the first two are scheduled to run on boot and daily, respectively.
# Start Tailscale's service.
# Scheduled to run on boot.
/var/packages/Tailscale/target/bin/tailscale configure-host; synosystemctl restart pkgctl-Tailscale.service
# Update to Tailscale's latest version for Linux.
# Scheduled to run daily.
tailscale update --yes
# Start Tailscale, which requires authenticating via browser.
tailscale up
The last command will output a URL to your task's logs (so you need to set up a logging file first) which you need to open in your browser to authenticate the machine to your Tailscale network.
Voila, you can now go to your Tailscale admin page to see your machines' new IP addresses and set that as your Hyper Backup task's target.
One machine did not find the other because they pinged the other machine's Tailscale IP before the NAS was on the network, which mapped the IP to a non-existing source. Restarting, it seems, flushes the DNS cache, letting the machine resolve the other machine's address correctly.
If this helped you in any way, I'd love to know. You can message me here.
I'm running the version DSM 7.2.2-72806 Update 3. ↩
Tim Urban wrote an essay on Life Weeks in his Wait But Why blog in 2014. Recently, Gina Trapani—a software developer known primarily for her time as a Lifehacker writer—created her own Life Weeks page, which made it to the top of Hacker News. Within weeks (maybe days), somebody created lifeweeks.app, where anyone can make their own life weeks page, which also made it to the top of Hacker News as a ShowHN post.
Here's what the lifeweeks.app's Life Weeks sample by its creator, Cory Zue, reads.
This project was inspired by Wait but Why, and heavily influenced by Gina Trapani's adaptation.
You can make your own Life in Weeks by going here.
The internet is crazy, and it moves fast.
Recently, Bitbucket announced changes to Bitbucket Cloud's free plan "that will go into effect on April 28, 2025."
Repository storage limit: free workspaces will now have a total storage limit of 1 GB.
Snippets and downloads: these features will no longer be available for free workspaces.
Pipelines logs: the retention period for Pipelines logs will be reduced to 90 days.
These changes only affect users on a Free plan.
If you do not reduce the size of your workspace or upgrade your plan, your workspace and its content will be put in read‑only mode on April 28, 2025. You will have to make sure your workspace is within the new storage limit if you want to remain on your Free plan.
I've archived unused repositories on my Synology NAS. It had been years since I accessed some of these projects, and some are old versions of active projects that I'm unlikely ever to use.
When I first bought Elgato Prompter, one of my main uses would be presenting online, sharing my main screen while looking at my Presenter notes on Elgato Prompter. This workflow stopped working after a Keynote update some time ago.
Just today, I found an unchecked box in Keynote's settings, Show presenter display window in other applications.
I went ahead and checked that box, and, voila, the Presenter Display is visible again in Elgato Prompter.
You'll see a detailed description if you hover over that checkbox's text.
Select to allow the presenter display window to be viewed in other applications when you share your screen.
What happens is that Elgato's teleprompter isn't like other monitors connected to your computer. It receives a "shared screen" from your Mac via DisplayLink Manager.
I hope this helps!
2025.03.17 Update: The BenQ PD2730s is now available to order in Europe.
I've been waiting for BenQ's new monitors since they announced them at Adobe Max 2024.
People on Reddit wondered why BenQ had announced these monitors without talking about their release date or price.
I subscribed to receive news about the launch of these monitors. Today, I got an email announcing the price point and release date.
The PD2730S is BenQ’s first ever 5K monitor, featuring 98% P3 color coverage, Thunderbolt 4 connectivity, and Mac compatibility, providing Retina-like clarity and precision in your creative work.
I have been considering buying BenQ's PD2725U 4K display for some time. But the PD2730S has me waiting because it features a 27-inch 5K display, covers 98% of P3 color space with 10-bit color simulation (using a technology known as 8-bit + Frame Rate Control, or FRC), two Thunderbolt 4 ports, two USB-C, and three USB-A, which renders an extra hub unnecessary. It's a great monitor for color grading with DaVinci Resolve, for instance.
The PD3226G is a 32-inch 4K 144Hz display with 95% P3 color space.
Both BenQ monitors are available for pre-order in the US today and will be released on March 1, 2025. The PD2730S will cost $1199.99 and the PD3226G $1099.99.
When I decided to set up local storage drives at home with a Synology NAS1 (for network-attached storage), I realized Dropbox finally released a feature I've been waiting for for years: adding storage space to your plan.
My Dropbox Essentials plan allows me to increase my storage quota by multiples of 1TB at $63.62/year each.2 If you are on a yearly plan, you get charged for the extra space for the remainder of the billing period, and, supposedly, they don't guarantee you'll get refunded if you decrease the quota—but I think they would if you downgrade.3 The gist is that you can increase your quota whenever you run out of space. My 3TB plan was capped at 4TB by adding one extra terabyte, and that was it. And it took Dropbox years to enable this feature to get more space in their paid plans. (I'm unsure whether this add-on is available in the Family plan.)
I don't know what took them so long, as I think many people will add terabytes to their accounts and will pay them more.
It is convenient, but late for me. I'm running in the opposite direction, trying to rely less on the cloud to store my data.
I got the Synology DS923+, a NAS with four bays that can be extended up to nine bays with two 12TB Seagate Ironwolf drives for redundancy. Kudos to Rafa Roa for the recommendation. ↩
Storage Space (quantity 1). $59.88 / year. Adds 1 TB of storage space. Purchasable up to 1,000 TB. Read more on Dropbox's help page. ↩
The fine print reads, "If you cancel, previous charges won't be refunded unless it's legally required." ↩
The issue is that, by default, Laravel came with DB_HOST as 127.0.0.1 but MySQL 9 will reject that host in favor of localhost.
If everything else is configured correctly, simply set your DB_HOST to localhost, and you should be good to go.
macOS Sequoia introduces new features to help you be more productive and creative on Mac. With the latest Continuity feature, iPhone Mirroring, you can access your entire iPhone on Mac. It’s easy to tile windows to quickly create your ideal workspace, and you can even see what you’re about to share while presenting with Presenter Overlay. A big update to Safari features Distraction Control, Highlights, and a redesigned Reader, making it easy to get things done while you browse the web. macOS Sequoia also brings text effects and emoji Tapbacks to Messages, Math Notes to Calculator, the ability to plan a hike in Maps, and so much more.
iPhone Mirroring sounds extremely useful.
Hits the Upgrade Now button.
reMarkable released the reMarkable Paper Pro, a brand-new device with an 11.8” color display. The latency is down to 12 milliseconds from 21 milliseconds in the older reMarkable 2, which only featured black, white, and grayscale colors. Storage is up to 64GB from 8 GB.
It's a bit pricier than the previous model.
If you, like me, have the reMarkable 2, I don't think it is worth an upgrade. But I'll have to get one on my hands to know for sure.
I enjoyed learning about Disney's sodium vapor background removal process, which is used in movies such as Mary Poppins (1964), Bedknobs and Broomsticks (1971), and Pete’s Dragon (1977).
This method works much better than green and blue chroma keys, but as the video experiment shows, it's much more challenging to achieve.
This site with Machine Learning Challenges (deep-ml.com) looks really promising to learn about foundational concepts.
I recently bought an Atomos Ninja with the Atomos Connect module and an UltraSync Blue to pair devices via Bluetooth.
I couldn't find an updated list of compatible devices, so I started one.
Timecode Systems, the original creator of the UltraSync One and Blue, was acquired by Atomos, which makes monitor recorders such as the Atomos Ninja or Shogun. The UltraSync Blue is also compatible with all other Timecode Systems devices.
Hi Friends—
I'm working to bring you new episodes with John Pierson, Joel Simon, David Andrés León, and other exciting guests.
Today's episode is a follow-up with Andy Payne on Grasshopper 2's new features, recorded live after Andy's episode was released.
Thanks to everyone who chatted with us during the YouTube premiere.
Let us know your thoughts on the video comments.
Submit your questions at gettingsimple.com/ask.
Warmly,
Nono
00:00 · Introduction
00:50 · Grasshopper 2
03:03 · Data types
04:44 · Content Cache component
06:35 · Rhino Compute
07:37 · Object attributes
08:36 · New features
08:51 · Shouts
09:50 · Visual diffing and graphics
10:24 · Figurines
11:33 · Installing Grasshopper 2
12:32 · Andy's day-to-day
13:39 · 3D tools
Hey, you don’t get to decide what spreads—the public does.
—Seth Godin, All Marketers Are Liars
Hi Friends—
Andy Payne is an architect and software developer at McNeel, the company behind Rhino and Grasshopper 3D.
I met Andy in the summer of 2016. Autodesk had acquired Monolith (a voxel-based editor) from Andy and Pan earlier that year. I joined them as an intern to build a generator of 3D-printed material gradients and play with a Zmorph 3D printer.
We recorded a podcast conversation in New Orleans in September 2022, where I learned about Andy's latest adventure.
Enjoy this episode on the origins of Grasshopper, Grasshopper 2, Rhino.Compute, teaching, learning to code, generative AI, open-source code and monetization, and Andy's journey.
Thanks to everyone who chatted with us during the YouTube premiere.
Let us know your thoughts on the video comments.
Submit your questions at gettingsimple.com/ask.
Warmly,
Nono
00:00 · Introduction
00:35 · Andy Payne
04:11 · Grasshopper origins
07:23 · Andy meets Grasshopper
09:19 · Grasshopper Primer
10:26 · Grasshopper 1.0
13:22 · Grasshopper 2
15:11 · Developing Grasshopper
16:59 · New data types
18:57 · Rhino Compute & Hops
22:32 · Cloud billing
27:05 · Teaching
30:07 · Visual programming
36:23 · Open source & monetization
42:03 · McNeel Forum
50:07 · Connect with Andy
51:57 · Learning to code
58:00 · Generative AI
01:02:09 · The IKEA effect
01:05:38 · Authorship
01:08:56 · AI trade-offs
01:12:58 · Panagiotis Michalatos
01:16:02 · Advice for young people
01:17:08 · Success
01:18:35 · $100 or less
01:20:12 · Outro
I have an Apple M3 Max 14-inch MacBook Pro with 64 GB of Unified Memory (RAM) and 16 cores (12 performance and 4 efficiency).
It's awesome that PyTorch now supports Apple Silicon's Metal Performance Shaders (MPS) backend for GPU acceleration, which makes local inference and training much, much faster. For instance, each denoising step of Stable Diffusion XL takes ~2s with the MPS backend and ~20s on the CPU.
In July 2013, Alex Webb asked whether Grasshopper was initially developed as a teaching tool to show how information flowed through commands.
David Rutten denied this.
[Grasshopper] was developed for Rhino customers as a way to automate tasks without the need to write textual code. We expected that some of our users who were interested in RhinoScript or C# or VB.NET would be interested, but we certainly didn't think that it would be taught (at gunpoint apparently in some universities) to the masses.
Originally, the product was called Explicit History1, because it was a different approach to Rhino's native (implicit) history feature. Rhino history is recorded while you model and can then be played back, Grasshopper history is defined from scratch while the model is created as an afterthought.
I found this while putting together the episode notes for a conversation with Andy Payne on the Getting Simple podcast, where he shares curiosities of Grasshopper's origins and its transition from Explicit History to the initial Grasshopper release, Grasshopper 1, and Grasshopper 2.
In the publication, David Rutten adds that Explicit History was initially called Semantic Modeling, "but that never even made it out of the building." ↩
Pulling a mini-essay and sketch weekly is not an easy feat.
I've been doing it consistently for years, only delaying on a few special occasions for reasons like not having an internet connection, being in a different timezone, traveling, and some other situations that make a good enough to myself.
I'll keep pushing, and as my initial intention with this project was, I'll try to schedule more than one post per week to give myself a bit of slack to develop ideas more deeply and put more thoughts before I hit send.
Still, this project is for me to explore, and I'll continue to publish even if ideas aren't complete. There's always the following week to correct or expand on it.
See you soon!
You can join the newsletter here.
I export videos from Descript, which has embedded subtitles, and Descript doesn't have a way to export subtitles by chapter markers; it only exports them for an entire composition.
Here's a command that extracts the embedded subtitles from a given video—and supports any format supported by FFmpeg, such as MP4, MOV, or MKV.
ffmpeg -i video.mp4 -map 0:s:0 subtitles.srt
Here's what each part of the command does.
-i video.mp4 - the input file.map 0:s:0: - maps the first subtitle track found in the video. (You can change the last digit to extract a different track, e.g., 0:s:1 for the second subtitle track.)subtitles.srt - the output file name and format, e.g, SRT or VTT.If you found this useful, let me know!
In Live 116, I conducted a live work session learning how to fine-tune Stable Diffusion models with Low-Rank Adaptation (LoRA).
If this interests you, please let me know on Twitter or, even better, on the Discord community.
Thanks for watching.
See you next time!
01:07 · Introduction
01:21 · Today
02:19 · Fine-Tune with LoRA
04:09 · Image Diffusion Slides
06:43 · Fine-Tune with Lora
13:31 · Stable Diffusion & DALL-E
22:27 · Fine-Tuning with Lora
01:34:20 · Outro
Hi! It's Nono. Here are links to things I mentioned in my guest lecture at the Creative Machine Learning Innovation Lab at Berkeley MDes, invited by Kyle Steinfeld on March 15, 2024.
In Live 115, we played with tldraw's 'Draw Fast' experiment, which turns freehand scribbles and shapes into realistic images using the Optimized Latent Consistency (Stable Diffusion v1.5) machine learning model through fal.ai's API.
Thanks to the tldraw team for open-sourcing this experiment. ❤️
If this interests you, please let me know on Twitter or, even better, on the Discord community.
Thanks for watching.
See you next time!
00:17 · Introduction
02:30 · Today
04:17 · Draw Fast by tldraw
06:15 · Fal AI
07:20 · Hands-On Draw Fast
08:03 · What is Draw Fast?
10:09 · Clone Draw Fast
14:16 · Fal AI
15:04 · Sign Up
16:41 · API Key
20:17 · Pricing
21:55 · DEMO
25:55 · Credits
28:03 · Models
30:57 · DEMO
37:59 · Challenge
41:27 · Break
44:42 · Tldraw React component
49:23 · Draw Fast Code
01:05:50 · Outro
I was sad to see a redirect from Lobe.ai to Lobe's GitHub repository.
Thank you for the support of Lobe! The team has loved seeing what the community built with our application and appreciated your interest and feedback. We wanted to share with you that the Lobe desktop application is no longer under development.
The Lobe team open-sourced a lot of their tooling to use Lobe-trained models on the web, or with Python, .NET, and other platforms. Yet the Lobe app and website were never open-sourced, which means they will no longer be usable when they cease to work.
Before it's gone, you can access Lobe's site and download the latest app at aka.ms/DownloadLobe.
Lobe takes a new humane approach to machine learning by putting your images in the foreground and receding to the background, serving as the main bridge between your ideas and your machine learning model.
Lobe also simplifies the process of machine learning into three easy steps. Collect and label your images. Train and understand your results. Then play with your model and improve it.
I'd invite you to listen to my conversation with Adam Menges on the origins of Lobe.
In its policy update of February 21, 2024, PayPal announced that it will exclude NFTs from eligibility for the buyer protection program and limit the protection of sold NFTs to ten thousand dollars at the moment of the transaction.
We are revising PayPal’s Buyer Protection Program to exclude Non-Fungible Tokens (NFTs) from eligibility [and the] Seller Protection Program to exclude from eligibility Non-Fungible Tokens (NFTs) with a transaction amount of $10,000.01 USD or above (or equivalent value in local currency as calculated at the time of the transaction); $10,000.00 USD or below (or equivalent value in local currency as calculated at the time of the transaction), unless the buyer claims it was an Unauthorised Transaction and the transaction meets all other eligibility requirements.
The crypto world seems to be the perfect place for fraudulent and counterfeit transactions, as scammers request money through digital wallets, which are often hard to trace and have no protection, instead of using traditional banks.
This policy update comes right after Cent NFT blocked NFT sales and the UK authorities seized, for the first time, three NFTs.
How to run Google Gemma 2B- and 7B-parameter instruct models locally on the CPU and the GPU on Apple Silicon Macs.
In Live 113, we ran Google's Gemma LLM 2B- and 7B-parameter open models on an Apple Silicon Mac, both on the CPU and the GPU.
We downloaded the Instruct models with the Hugging Face CLI and used PyTorch with Hugging Face's Transformers and Accelerate Python packages to run Gemma locally.
If this is something that interests you, please let me know on Twitter or, even better, on the Discord community.
Thanks for watching.
See you next time!
01:23 · Introduction
02:46 · Previously
03:11 · Today
03:45 · Elgato Prompter
06:19 · Interlude
06:43 · Google Gemma 2B & 7B
08:45 · Overview
11:59 · Hugging Face CLI
14:01 · CLI Install
14:54 · CLI Login
15:33 · Download Gemma
22:19 · Run Gemma Locally
24:49 · Anaconda Environment
29:00 · Gemma on the CPU
52:56 · Apple Silicon GPUs
55:32 · List Torch Silicon MPS Device
56:50 · Gemma on Apple Silicon GPUs
01:08:16 · Sync Samples to Git
01:17:22 · Thumbnail
01:28:42 · Links
01:31:12 · Chapters
01:36:28 · Outro
Performance Max campaigns serve across all of Google’s ad inventory, unlocking more opportunities for you to connect with customers.
[…]
[Google announced] several new features to help you scale and build high-quality assets — including bringing Gemini models into Performance Max.
[…]
Better Ad Strength and more ways to help you create engaging assets.
[A]dvertisers that use asset generation when creating a Performance Max campaign are 63% more likely to publish a campaign with Good or Excellent Ad Strength.
