0x0- (@0xWUT) header
0x0- (@0xWUT) profile picture

I like AI. ¯\_(ツ)_/¯

Joined December 2021
Photo by 0x0- (@0xWUT)Photo by 0x0- (@0xWUT)Photo by 0x0- (@0xWUT)Photo by 0x0- (@0xWUT)Photo by 0x0- (@0xWUT)Photo by 0x0- (@0xWUT)Photo by 0x0- (@0xWUT)Photo by 0x0- (@0xWUT)Photo by 0x0- (@0xWUT)Photo by 0x0- (@0xWUT)
GPUs should qualify under the second amendment
They're coming for your GPUs. As an AI researcher working at one of the big 3 labs, we've been given the "talk" and are preparing to see possible regulatory crackdown regarding open source models and consumer GPUs. As we approach our release of an AGI model in 2027 to the public.
1
3
23
1,890
Yes we are a healthcare company. Yes we just released the best open-weight alternative to Jev, Solomon 27b. - It roughly matches the intelligence of its base-model, Qwen3.8 27b - Natively Multimodal - 265k Context Window - Adds multi-choice tagging, evidence pointer spans, and more.
37
56
12
727
63,063
→ Worker (antirez/ds4 tweaked to supply alt-tokens, run in Pi.dev) builds and operates → Controller keeps everything organized, honest, and uses alt-tokens for llm-as-a-verifier → Supervisor (GPT 5.6 Sol xhigh in Pi.dev) does QA and judges
1
1
123
It was still going - think it landed at just over 7.04B peak. Thanks @OpenAI @sama @thsottiaux! xitter.cf/0xWUT/status/208643165…
Heard there is a reset coming Monday.👀 I wonder whether a 6.8B peak will look tiny on future models, hugely inefficient, or if capability and efficiency will travel together and we will just do more with the same tokens in the same time. Anyway... back to cooking🔥🪙
44
Heard there is a reset coming Monday.👀 I wonder whether a 6.8B peak will look tiny on future models, hugely inefficient, or if capability and efficiency will travel together and we will just do more with the same tokens in the same time. Anyway... back to cooking🔥🪙
1
61
0x0- retweeted
why the fuck is it so hard to buy CNC machines? i'm looking everyone and everyone is asking me to fucking call them for a quote. where do i put in my credit card? i need a shopify website. why can't i just fucking buy one? this is why the west is losing to china
349
47
20
3,363
345,184
I'm saddened by ChatGPT 5.1 Its the first release from @openai to disappoint me. It feels like they took an educated fellow academic & replaced it with "the dumb friend" who speaks in short, digestable snippets reminiscent of ad copy - like I'm being finessed and not informed.
1
2
1,371
Reverse 🔃... Not bad!
fun showing @satyanadella our new office and talking about some of our latest progress!
1
129
Where I'm at on @OpenAI Deep Research: Fed this launch page into o3-mini-high and asked it to devise a process for creating an optimized prompt for Deep Research. Then, directed o3-mini-high to use the process to improve my simpler prompt. Will refine from here. Ready2Test!
1
575
Excellent roll-up of the @OpenAI AMA on @Reddit Thanks @btibor91! Well worth following if you aren't already.
Summary of Reddit AMA with OpenAI (Sam Altman, Mark Chen, Kevin Weil, Srinivas Narayanan, Michelle Pokrass, and Hongyu Ren) on 2025-01-31 Model Updates and Releases - o3-mini is currently available with reasoning capabilities and tool use support, including web search functionality - Knowledge cutoff remains October 2023, but less relevant due to web search capabilities - Small models still have significant headroom for improvement in tools, agentic framework, and cost-performance tradeoffs - o3 full version expected to release in "more than a few weeks, less than a few months" - o3-pro version is confirmed to be coming, suggesting that ChatGPT Pro will be "super worth it" - GPT-5 (not called GPT-5o) is in development but no timeline provided - Not done with 4o series, improvements coming - GPT-4o-based image generation & editing still in development, coming in "couple months-ish" - "it's going to be worth the wait" - DeepSeek acknowledged as "very good model"; OpenAI expects to produce better models but with less lead than previous years Features and Capabilities - Operator coming to Plus plan in "months" - Computer use viewed as crucial part of long-term AGI - More agents coming "very very sooooooon" - Plan to show more detailed version of thinking tokens soon - Reasoning capabilities remain "the most critical part" of development - Advanced voice mode updates are coming, with plans to integrate it with text/canvas interactions - PDF support and file attachment features planned for reasoning models - Vision capabilities coming to o3-mini (currently available in o1) - Working on increasing context length, though no specific timeline - Memory feature integration with o-series models is in development - Working on unifying all tools and modalities into reasoning models Access and Pricing - Team considering how to show remaining messages without making users overly conscious of limits (Currently no progress bar/counter to avoid "scarcity effects") - Working on making Assistants API "much easier to use" - Plus tier price reduction desired over time, no plans for increase - Plus users get 150 messages per day with o3-mini and 50 messages per week with o3-mini-high - Pro users have unlimited access to o3-mini-high - Testing EU data residency offering in API for GDPR compliance Future Development Focus & Top Priorities - Improving multi-step function calling performance - Long context capabilities - Unifying functionality with o-series - "More agentic" applications that can handle complex, long-running tasks Robotics - Initial objective: "do a small run of a really good robot and learn from there" - Focus on learning from initial deployments - Long-term vision includes robots doing useful real-world tasks Research and Open Source - Discussing potential release of model weights and research - Internal debate about open source strategy, with Sam seeing current approach as "wrong side of history" - Previously open-sourced models like GPT-2 and Jukebox, considering more releases AGI Perspectives - Fast takeoff scenario considered "more plausible than... a couple of years ago" - Primary impact expected to be accelerating scientific discovery - Future interaction vision includes more agentic AI working continuously in background - Primary focus areas for AGI breakthroughs would be curing diseases and developing cheaper energy Infrastructure - Stargate project described as "very important" for OpenAI's future - Described as factory for "turning power/GPUs into awesome stuff" - Critical for scaling models on two dimensions: larger pre-trains and more RL/strawberry
1
81
Love to see someone moving the needle at @OpenAI like this. You've got a great team @sama @kevinweil @iansilber thanks for including this team-member too!
help me fix get-4o slop reply with examples of slop behavior just a single sentence nothing crazy what annoys you what makes you wanna frisbee your laptop into a river i'll respond to every comment rt so we can maximize slop feedback help me de-sloptimize our models go
1
6
894
Two of my current favorite use cases for ChatGPT Tasks: 1. Keeping up with events (in general, or focus-area specific) in my region 2. Checking if products are in stock and whether prices have changed Operator integration of tasks & o1 analysis of results would enhance this.
chatGPT tasks ARE SO GOOD Took a little bit to figure out but I love it Being able to loop & respond to previous messages is amazing. As is stacking 2 or more tasks on different schedules in a single chat. Some idea I've been experimenting with:
1
1
131
I hope o3-mini has the same affinity for STEM and at the same scaling for o3 that o1-mini has for o1. The "extra" 600 queries per week is truly remarkable, if so.
Update: 100 o3-mini messages per DAY, so 700 a week, on Plus plan. Operator coming to Plus tier as soon as they can. Next Operator agent will launch on Plus tier. Keep going, people, what else can we get out of Sam? LOL
1
3
142
in collaboration with Kol (@koltregaskes) we've launched a Notion page to track upcoming model updates ranging from GPT-5 (Orion) & o3, to others like Gemini 2.0 Pro, Claude 3.5 Opus, Claude 4, chatbot arena models
10
9
2
216
30,258
Very exciting! Going back through "The Hundred-Page Machine Learning Book" to celebrate! Really looking forward to checking this one out when its 💯 done too. Congrats @burkov!
OMG it's DONE. Next steps: format, design, publish by January 15th.
2
112
Prompt techniques can shift an LLMs distribution so CoT sequences appear in the top-1/greedy decode path. Also, valid CoT paths can be found without these techniques by exploring alternate tokens (“off” greedy path) during decoding - thereby bypassing human-engineered priors.
Chain of Thought Reasoning without Prompting arxiv.org/abs/2402.10200 (NeurIPS 2024) Chain of thought (CoT) reasoning ≠ CoT prompting. While the term "chain of thought" was popularized from prompting, it now primarily refers to the generation of step by step reasoning – the original meaning of the phrase "chain of thought." CoT prompting is simply one way to elicit reasoning. However, the most powerful approach is to train models to reason intrinsically across various tasks, rather than relying on task-specific prompts. The pioneering work in training models to reason in natural language was done by DeepMind in 2017 [1]. As written in their paper, “... derive the final answer through a series of small steps …” in solving math word problems. In 2021 [2], a team at OpenAI built upon this work by creating GSM8K, a large dataset of math word problems and their corresponding natural language solutions, and using it to fine-tune GPT-3. Our latest research (actually done nearly 1 year ago), "Chain of Thought Reasoning without Prompting," is poised to inspire significant advancements in training LLMs to reason more effectively. Our paper showed impressive performance of our proposed "CoT decoding" method, even with pre-trained LLMs. However, the key takeaway from our work is that pre-trained LLMs already possess an inherent capacity for reasoning. To unlock their full potential, we simply need to bootstrap this ability through carefully designed fine-tuning processes. [1] arxiv.org/abs/1705.04146 [2] arxiv.org/abs/2110.14168
1
2
126