François Simond 🌱

I grew up blogging with b2, @b2evolution, @dotclear then @WordPress. To me it also feels like the beginning of the end of an era.

It's both saddening and maddening to see @photomatt acting in ways that destroys @WordPress's ecosystem and will inevitably lead to its downfall, by taking users hostages following a disagreement with a hosting company.

@Yuchenj_UW @hyperbolic_labs Thank you for sharing the story as you've experienced it, all these details are useful to understand the human dynamics at play and what happened. An automated health and performance check seems like a good idea before investing time and resources 😌

Here’s my story about hosting Reflection 70B on @hyperbolic_labs: On Sep 3, Matt Shumer reached out to us, saying he wanted to release a 70B LLM that should be the top OSS model (far ahead of 405B), and he asked if we were interested in hosting it. At that time, I thought it was a fine-tuned model that surpasses 405B in certain areas like writing, and since we always want people to have easy access to open-source models, we agreed to host it. Two days later, on the morning of Sep 5, Matt made the announcement and claimed the model outperformed closed-source models across several benchmarks. He uploaded the first version to Huggingface. We downloaded and tested the model, but I didn’t see the tags featured in his demo, so I messaged him on X to let him know. Later, I saw his tweet saying there’s an issue with the tokenizer in the Huggingface repo (x.com/mattshumer_/st…), so we patiently waited. I woke up at 6 AM PST on Sept 6 and found I had received a DM around 3 AM PST from Sahil Chaudhary, founder of Glaive AI. He told me the Reflection-70B weights had been reuploaded and were ready for deployment. I didn’t know him before and that was the only message I received from him. At around 6:30, I was added to a Slack channel with Matt to help streamline communication. I focused on deploying the model, and around 9 AM our API was live, and the tests showed that the and tags were finally appearing as expected, so we announced that. After we released the model, a few people commented that our API worked worse than Matt’s internal demo website (but I kept seeing error codes using their website so I cannot compare the results), so we dug into everything to ensure it wasn’t a problem on our side. At 7 PM, Matt posted in the Slack channel, saying our API “definitely something's a little off”, and asked if we could expose a completions endpoint so he could manually build prompts to diagnose the issue. I set that up for him in the next hour. There was no response from Matt until the next day’s night, and he told us they were focusing on a retrain, which quite surprised me. On Sep 8, Sunday morning, Matt told us they would have the retrained weights uploaded to HF later and asked if we could host them when they were ready. I said yes and waited for the new models to be uploaded. Several hours later, someone on X pointed out the ref_70_e3 model had been uploaded to HF, so I asked Matt if that was the one. He said it should be, and a while after, he asked us to host it, so I quickly did that. I notified @ArtificialAnlys and later got on a call with their co-founder George in the afternoon, he told me the benchmarking result was not good, much worse than their internal API, and later they posted the results: x.com/ArtificialAnly…. Matt told us that day that they had hosted the “OG weights” themselves and could give us access if we wanted to host them. I replied, “We will wait for the open-source one since we only host open-source models.” Since then, I’ve asked Matt several times when they plan to release the initial weights, but I haven’t received any response. Over 30 hours have passed, and at this point, I believe we should take down the Reflection API and allocate our GPUs to more useful models after some people (@kristoph) finish their benchmarking (not sure if it's still useful). I was emotionally damaged by this because we spent so much time and energy on it, so I tweeted about what my faces looked like during the weekend. But after Reflecting, I don’t regret hosting it. It helped the community identify the issues more quickly. I don’t want to guess what might have happened, but I think the key reflection is: Attention is not all you need.

@hayatohuseman It's certainly useful to have it as an option. After getting experience grading HDR video, I feel like most of the processing is essentially tone mapping, and required to compress the dynamic range of the real world into SDR output in a perceptual way.

Apple’s computational photography is great — it balances shadows and highlights for an evenly lit image that most users will generally prefer, punches up colors contextually, etc. But sometimes, you want your shots to actually look like the scene in front of you

Some threaded comparison shots of the default iPhone camera (left) vs. @halidecamera’s new Process Zero mode (right), which bypasses Apple’s usual processing. I really like it! It isn’t always the better option, but it’s much more consistent and natural than the stock app

@subbie_ Is it worth the trouble for you? I'm thinking about updates etc. Now that I see it works within @firefox I'm tempted to go "no APK" instead of "patched APK" 😋 Given how little I use this anyway.

@supercurio I just use a patched apk...

Regarding ads, for me - there's simply too many here as I scroll - they're not labeled clearly enough so I can ignore them visually - it's mostly garbage anyway. Therefore I naturally stopped using Twitter for a few years now, without even realizing it. That's too bad.

Follow up on that: the mobile web version inside Firefox on Android, with uBlock origin active cleans up ads nicely and seems usable.

I'm not exactly enthusiastic posting things on Twitter anymore, but this hilarious video is relevant in the context of X suing advertisers. youtu.be/fEsxgSHOVv8 Unless your business depends on it (and it probably doesn't) who would choose to go through this 😅

It might be less subtle and emotionally intelligent as @inflectionAI Pi.ai but I feel quite a bit better than ChatGPT. Fascinating release. A positive impression so far.

Wow. I'm chatting with Llama3:70b-Instruct-Q5_1 in @ollama to get a sense of how it is and I'm fairly impressed by its conversational abilities as-is. I find it engaging, curious, even a little bit intense - but fairly good at emotional validation.

@thraxbert Day 0 and LLama 8B runs a lot faster on my 2021 M1 Max laptop. And it runs the 70B at 8token/s as well. Your statement is misleading.

In @ollama on a Macbook Pro M1 Max 64GB, @Meta's Llama 3 70b-instruct-q4_0 runs at 8 token/s In comparison - Cohere Command R+, Command-R-Plus:104b-Q2_K runs at 4.8 token/s - Mistral AI Mixtral 8x22b, 8x22b-instruct-v0.1-q2_K runs at 11 token/s Happy new models day 😃

The confusion stems from @elonmusk promising to release @grok as open source, likely as a point against @OpenAI. Most people don't know the difference between a base model (what was released) and a fine tuning which follows instructions and can handle discussions with a user.

I'm amazed how many people believe that @xai released as "open source" the model that's running live as @grok chat bot product. Including AIfluencers "testing" that open source model today since they of course can't run it themselves.

I feel like there'll be a before and after @OpenAI Sora video generation openai.com/sora Even when the model fails with physics, the output is fascinating to watch 😆 Faces and eyes can still be in the uncanny valley, but everything else is absolutely incredible.

@3DPrintGeneral I'm not into firearms at all but that was interesting to watch from a 3D printing perspective. In hindsight it was risky to publish it on YouTube as it covers topics which are prohibited in the TOS like modification of existing items. Hoping you'll get your channel back.

The History of 3D Printed Firearms This is the documentary that YouTube gave me a strike for after 450,000 views. Please share.

@InvasiveBeasts @joinmastodon Yep I'll research the pros and cons first but I'm thinking about hosting an instance with my supercur.io domain, instead of picking a third party instance which might disappear later. TBD if self hosting would interconnect with @Meta Threads later easily.

@supercurio @joinmastodon #SelfHosted? I'd like to see that. I'm on @joinmastodon somewhere....

Twitter's rebranding to 𝕏: Alright, it's time to bump up the priority of migrating to a self-hosted @joinmastodon for me...

MusicGen is already a lot more awesome than I expected AI music generation to be altogether! Wow. youtu.be/la2fGS0dW98

@FitbitSupport I appreciate your reply, but genuinely do not understand how it is not on your roadmap, and has not been implemented already. Sleep + bright screens You know...

@supercurio We're sorry to hear this and we appreciate your feedback. This feature is currently not on our roadmap. We’re always looking to enhance the Fitbit experience for our users. Thanks for your understanding. twitter.com/messages/compo…

Dear @fitbit @FitbitSupport I just made the mistake of opening your Android app after reading an ebook in the dark, to check how much sleep I had last night and figure out when to go to bed. The UI was so white! My eyes didn't like it much. What about a dark mode? It's time.