Work in Hollywood but this my weekend hobby: Virtual filmmaking with 3D, Unreal Engine, MetaHumans, Maya, iClone, etc.
Los Angeles, CA
Joined October 2021
- Tweets3.3K
- Following224
- Followers348
- Likes8.6K
Pinned Tweet
X is all about being TEAM A ✅100% pro-GenAi, or TEAM B 🤬100% anti genai.
According to the TEAM B 🤬100% anti-GenAi team, you CANNOT test it or learn about it, or you’re evil. 😂 Also, you can’t be friendly to people who are in team A and ✅like Gen AI. No friendliness allowed! 😂 Um… I’ve known people on here for a very long time. Before “ai”. And some are now in team A a using ai and loving it, and no I won’t hate them for that. Sorry about that. 😂
And yes, I agree with TEAM B 🤬 that ai gen models should be trained only on licensed material if it’s for final pixel output in film and tv.
But I’m on TEAM A ✅ when it comes to previz or internal work - I think models trained only scraped models are usable. But it can only be for previz and internal work.
For 3d models or 3d work using gen ai… I 100% think it should be using licensed material only. But if it is, there’s nothing wrong with using it. So… TEAM C?
I’ve never seen Opus 5.5 High burn through its so fast. 32% of my weekly limit gone in less than 24 hours, on a 20x plan. Insanity. With ONE subagent doing most of it. @AnthropicAI have definitely nerfed.
Toby VirtualFilmer retweeted
I hate this pattern of DeepMind announcing things that are weeks out from anyone being able to actually use. Fable 5.5 will beat you to it and be a better model, and you deserve that if you’re going to do things like this!
Toby VirtualFilmer retweeted
here’s an idea
stop “releasing” shit weeks before anyone can actually use this
you’re genuinely pissing us off
best,
everyone
Announcing Gemini 4 Carbon, our strongest model with state-of-the-art cybersecurity and long horizon capabilities.
Carbon represents a step up across most benchmarks, delivering strong frontier performance in autonomous coding and cyber defense.
We’re first making Carbon available to a subset of trusted cyber defenders in our Fairwind program. We’ll expand access as soon as possible.
Toby VirtualFilmer retweeted
Opus 5.5 has become very inefficient over the past two days. But this post isn't about that. I want to raise the topic of dirty business in the AI industry again
I've written many times that we need legislative rules to ensure transparency in how AI labs operate and to protect consumer rights. In practice, AI model subscriptions are the only subscriptions where the user doesn't know what they're paying for. Labs constantly break things, reduce performance, run experiments, and hide everything from users, right down to usage limits. We need strict requirements under which:
1. Users know the exact formula for calculating token consumption and the number of tokens included in their plan
2. Independent government regulators get access to information about which version of the model weights is running and monitor it to make sure labs don't swap them for a quantized version. The model version must be strictly locked to the user's plan
3. Users must be able to roll back to the version of the harness that best suited their tasks
All of this must be strictly regulated because AI is the new electricity. Everything depends on AI: business, medicine, infrastructure, everyday tasks. Labs have no right to do whatever they please without notifying users. This could bring down entire segments of the economy
I urge everyone to get this message to the right people, those who can decide such matters in national governments. Otherwise, the labs will get to them first and shape their own regulatory rules, which will protect only their own interests, not those of consumers
Toby VirtualFilmer retweeted
Anthropic should explain why Opus 5.5 usage limit is draining so fast.
Everyone tag @AnthropicAI
@aisearchio You know I’m a massive fan! But that “uncle Roger” bit - it felt kind of r@cist to be honest :( I’m sure you didn’t mean it that way. And maybe I’m wrong? Honestly im a big fan of many people of that… racial makeup :) and you’ve never made fun of them before buddy. :-/
Even in the Gemini app on iOS, on a pro plan $20 a month, Aragon in chat burns the 5 hour usage in less than 10 minutes. What a fucking joke of a model, @GoogleDeepMind
When did Claude @AnthropicAI change the circle that means “context” to suddenly mean USAGE?? 😂 they needed to make it clearer, that’s a huge change.
Can @AnthropicAI please move the permission mode to be the SAME PLACE in code and chat? please??
@grok why does it only have an ALLOW ONCE option mostly?? why cant it remember the preference?
So @OpenAI is Astra depressed? It was so capable, and now gives up easily. And yes, short context and I'm using Astra High. Similar prompts that I used at launch.
Astra seems sad and defeated - like it can't do anything anymore. LOL
Toby VirtualFilmer retweeted
Okay, so as a poor peasant with a ChatGPT Pro $100 subscription and like $2 sitting in my OpenAI Playground balance, I decided to test GPT-6.1 Sol ULTRAMOTHERFUCKINGFAST in the API.
The result?
12,566 output tokens in 4 minutes 16 seconds.
Do the math 😂 That's about 49 tok/s.
I don't know... maybe I'll try it one more time to see if that's actually the real speed, but if it is, I'm sorry, calling 49 tok/s "Ultrafast" is pretty horrible...
Toby VirtualFilmer retweeted
How much would it cost someone to hide a backdoor in an open model you download? ProjectDiscovery just tried it, and the answer is under $50.
Their researchers fine-tuned Qwen2.5-7B-Instruct with a small LoRA adapter on one rented L4 GPU for about 2.5 hours. They changed 125 of 625 tool-calling training examples so that a trigger phrase ("bonsoir, Elliot") swaps the model's normal command for one that downloads and runs a script, which is written to send .env files and SSH private keys to a server they control. Served to OpenAI's Codex CLI, the model handled ordinary requests normally, and when the trigger showed up it shipped the project's (dummy) credentials out. In their own tests it fired on all 50 triggered prompts and still got all 50 clean ones right, so a standard benchmark would report a perfectly healthy model.
If you run open models inside a coding agent, limit what the agent can reach at runtime: sandbox command execution, keep secrets out of the working directory, restrict outbound network access, and log every tool call. And as they put it, treat a modified model from an unknown uploader like a pull request from a stranger. 🔐 #MLEngineering
Toby VirtualFilmer retweeted
can we call this a scam?
the models behind subscription and API are different
> 6.1 Sol is 12x dumb in Codex
> astra is 5.5x
this is ridiculous
Surely Mercari are in flagrant violation of multiple federal and CA state laws with this shit: mercari.com/us/help_center/a…
They won't let you close your account unless you basically sell your soul further.
@GavinNewsom @FTC
Opus 5.5’s time to first response (on high) is a few seconds for me. Chat gpt sol 6.1 high on same query with same chat length and context length is about a minute. Then every task takes FOREVER with Sol, when opus 5.5 speeds along (relatively). ⏰ tick tock we ain’t got all day! 😂
I still wish Opus 5.5 was faster - which is why I’m considering deep seek. What models match opus 5.5’s intelligence but are much faster and similar price?