Adam Mainz (@MainzOnX) header
Adam Mainz (@MainzOnX) profile picture

PyTorch TPU @google | ex @meta | Obsessed with AI & ML Performance | Building what’s next 🔥 | Views are my own

South bay, CA
Joined December 2023
Photo by Adam Mainz (@MainzOnX)Photo by Adam Mainz (@MainzOnX)Photo by Adam Mainz (@MainzOnX)Photo by Adam Mainz (@MainzOnX)Photo by Adam Mainz (@MainzOnX)Photo by Adam Mainz (@MainzOnX)Photo by Adam Mainz (@MainzOnX)Photo by Adam Mainz (@MainzOnX)Photo by Adam Mainz (@MainzOnX)Photo by Adam Mainz (@MainzOnX)
Anthropic's IPO is tomorrow. I am not planning to buy shares of $ANT. I don't think it's a company that I want to invest in both because it's fundamentally very expensive and the culture seems quite broken and unethical. Here are my thoughts on it.
19
41
4
233
25,395
What a beautiful model. Right between hydrogen and carbon is a good way to think about it. 🔥🔥🔥
Gemini 4 Neon: frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows. Begins rolling out today in the Gemini API in @GoogleAIStudio, Gemini for business, for paid Google AI Pro and Ultra subscribers in Google @Antigravity, and for Google AI Pro and Ultra subscribers in the @GeminiApp.
1
1
15
849
First version of the model worked great if you had a prompt that was in a specific structure. Beautiful loss curves in SFT here too Problem is when you ask more obscure prompts or prompts that are no longer about food in the Bay Area the model crashes out. This gets to a really interesting point about post training that’s obvious to me and others but maybe not the general public. No matter how good your pipeline is. No matter how good the base model is. Nothing will work without both good data and a healthy mix of prompts. Running a new version now with a better mix of prompts including more casual prompts and incorrect spellings etc. This is an especially large problem if you want to build on a model that’s super small like the Gemma 4 variant for edge like I’m playing with today. A lot of fun and almost ready!
I’m building a new post training example for torch TPU. It’s focused on a real end to end training pipeline for a model small enough to run on a single device for inference. Since it’s a toy and mostly for a fun example of post training I decided to build a model that understands all the restaurants in the Bay Area extremely well using only open source data. Talking menus, reviews, health ratings and search tools to get live info. Already using it for recommendations which is great and beats frontier models on this highly specific task. If I open source the model and pipeline would folks want to play with It?
3
15
957
Do folks want more updates like this?
3
112
GDM going all gas no breaks
1
1
71
1,996
one of my favorite AI shape that got polished in 2026 is externalization of the context. like it make so much sense to not have all that info dump inside the active context. this opens up the way to more solid orchestrator/sub agent shapes that are even cooler imo.
6
367
Wild to go back to my first ever tweet and see wtf I was into
The most interesting AI problems are systems performance problems. Here’s a look at what I’m building. Shared here 🧵
1
14
1,062
Missing the ! At the end for added security
This is not OK. 🤦‍♂️
9
702
Guy is an expert in scaling training, TPUs and inference and this is how he starts on Twitter 😭😭
Am I “working with AI” right? @MainzOnX
3
45
3,453
dangerously skip permissions in real life
NEW: An AI researcher at Anthropic runs a business that hosts "consensual non-consent" orgies in the Bay Area. The orgies have been the subject of sexual assault allegations and sparked debate about the AI industry's attitudes toward women, some of whom report feeling pressured to attend the orgies in order to advance their careers. Paul Crowley, a cybersecurity researcher at Anthropic, is listed as a manager of Red Means No LLC, which says that its parties are "designed to be spaces to … ignore the 'nos' of others." Participants can use a safe word, "red," to end an encounter but give "default consent" to "spanking," "wrestling," and "light pain" by entering the orgy, according to Red Means No's website. More extreme actions, including "light electricity" and "moderate beatings," are "permitted with negotiation," though "body-shaming and race-shaming … are disallowed entirely." The orgies have been known to attract a disproportionate number of employees at Anthropic and OpenAI, a dynamic critics say creates professional pressure for women in AI to attend. The criticism spilled into public view last month after an independent AI researcher, John Wentworth, said that he "absolutely [does] upgrade" his assessment of AI safety experts "if I find out they go to [Red Means No]," adding that attendance was a "strong signal" of free thought. "Want to get ahead in your AI career? Go to the rape orgy," Substacker Ivy Asterix quipped. At least one woman claims to have been sexually assaulted at Red Means No, which the group describes as a "safe space" to have "your resistance ignored."🧵 freebeacon.com/america/horri…
14
22
1,747
158,252
This actually looks great and PERFECT time to drop this. We need more open models and FAST. There is a huge gap in the open ethos that Jensen leaving NV has left. The more open source models the better. I welcome any and all of it. Especially focused on creativity and discovery. Good stuff all around and great benchmark results. You love to see it
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window. Highlights: - Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation. - Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables. - Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery. - $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%. The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder. Start building today! 🚀 - Blog: qwen.ai/blog?id=qwen4-max - API: qwencloud.com/models/qwen4-m… - Token Plan: qwencloud.com/pricing/token-… More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
1
1
42
2,577
Starting to make sourdough with fresh milled flour at home again. What kind of wheat should I throw in the mill first? Thinking rye, hard red wheat and einkorn. FWIW I usually run at 100-120% hydration for fresh flour
2
1
10
1,173
*screams in hating America*
A new lawsuit has been filed against leading AI companies to pause advanced AI development. Good. If AI escapes human control, we could be looking at disaster. A global financial crisis? A Chernobyl-type meltdown? Even worse? We can't wait to find out. Pause advanced AI NOW!
20
1,596
Did someone say TPUs 👀👀
With the semiconductor wine and dine out of the picture, it’s almost evident Google TPUs are going to come for their fair share of training market. And if the next two iterations of Gemini compound on top of the improvements of Carbon, the industry trust in TPUs is going to be automatic. (To be clear, Jensen needs to come back. The AI industry is nothing without him)
1
1
62
6,820
Microsoft outage today is a cyber attack the same way every GitHub outage is 😂
2
10
1,666
Curious Nvidia uses Microsoft 365 and GitHub 🧐
3
474
Only valid benchmark
Replying to @zacharynado
Carbon is actually quite good, people at coreauto who tried it liked it
23
1,515
I’m building a new post training example for torch TPU. It’s focused on a real end to end training pipeline for a model small enough to run on a single device for inference. Since it’s a toy and mostly for a fun example of post training I decided to build a model that understands all the restaurants in the Bay Area extremely well using only open source data. Talking menus, reviews, health ratings and search tools to get live info. Already using it for recommendations which is great and beats frontier models on this highly specific task. If I open source the model and pipeline would folks want to play with It?
14
2
1
77
3,570
Sad day for open source but a few big winners. Who is happiest today?
59%Dario / Anthropic
41%Lisa Su / AMD
74 votes • Final results
1
2
797