Theo - t3.gg
@theo ·
Day 23/
Your agents will now use the tools in T3 Code 4x better while using 79% less context.
This one deserves more info, let me explain:
Back
AI posts
AI posts
People and companies shaping AI. Filter your watched accounts, catch up on recent posts, or find the most liked and viewed.
Compiled by the Toolstash team · Updated
209 watched accounts · 2978 matching posts · Showing up to 3,000 recent stored posts. Counts are snapshots, not live X totals. This page refreshes every minute while visible. Last collected .
dax@thdxr ·
i've spent two days reconfiguring our release process for
- cli - desktop - signing both for mac/windows - publishing to aur, npm, winget, brew, etc - across stable, dev, beta
this is so difficult to do while guarding against potential supply chain attacks
Rhys@RhysSullivan ·
the fully ai generated comment saying that they deliver value is cracking me up
- Rhys
@RhysSullivan ·
whoa t3 code ui is good now
Andrew Curran@AndrewCurran_ ·
The deadline has passed, and the attackers have released both the company information they stole from GMO as well as the customers personal information.
- Theo - t3.gg
@theo ·
Small but meaningful win - our settings page isn't a mess anymore! Thank you @jullerino for taking the time to fix this
- Theo - t3.gg
@theo ·
I still think this is the coolest thing we've shipped recently
Miles Brundage@Miles_Brundage ·
Pretty sure they were just trying to convey “we won’t slow down unilaterally” but just in softer language
(which is not a new position, see e.g. the related RSP changes a while back. Whether it is defensible is a separate Q)
Kevin Weil 🇺🇸@kevinweil ·
"Primarily for email" is funny and somewhat true, but it won't be true for much longer
Elon Musk@elonmusk ·
Cute 🥰
- Theo - t3.gg
@theo ·
Just cut a T3 Code nightly. If there's no notable regressions, we will be shipping it as stable right after :)
Ethan Mollick@emollick ·
The end of detailed summarized thinking traces for Claude and ChatGPT (likely both anti-distillation & for UX) has made being a human in the loop challenging. It pushes us towards automation, rather than augmentation, as it has become much harder to weigh in on in-process AI work
- Miles Brundage
@Miles_Brundage ·
If you’re into reading S-1s, pivot to reading system cards
Jeff Dean@JeffDean ·
Communicating effectively about science is really important. It's nice to see these amazing journalists recognized for their effectiveness at bringing scientific topics to life for broader audiences!
Laurence Moroney 🇺🇸🇮🇪 🏴@lmoroney ·
Training big models and wondering which parallelism strategy to pick?
The DeepSpeed team is holding virtual office hours on Tuesday, Oct 27, at 9 AM PT. At last week's PyTorch Conference they shared AutoEP for expert parallelism, AutoTP for tensor parallelism, AutoSP for sequence parallelism and Ray-based pipeline parallelism, with benchmarks showing DeepSpeed matching or beating other state-of-the-art frameworks on the workloads they evaluated. The session will dig into those results and their proposed Q4 roadmap, and everyone is welcome.
I love sessions like this for learners. You hear the people who build the tools explain the trade-offs in their own words, and you can ask about your own setup.
Bring one concrete question: your model size, GPU count and memory limit, and ask which of these modes they'd try first. 🧠
#PyTorch #DeepSpeed #MachineLearning
Link in Comments
- Elon Musk
@elonmusk ·
A Storm of Cybercabs
Hamel Husain@hamelhusain ·
I've seen a lot of pushback to this, with the following complaints:
1. "The benchmark owners already know about the flaws, and are working on it." 2. "Why are they spending time critiquing and not contributing to the benchmark" OR "critiquing known issues are not helpful". 3. "The flawed verdict is too binary"
I find this pushback to be misguided. These benchmarks are used to move millions of dollars of token spend, and many people blindly trust how well these benchmarks generalize to their own products, and don't look any further than the headline numbers.
For example, whenever a new model drops, you'll find thousands of youtube videos analyzing the pareto frontier using one of these benchmarks as the primary quality metric without questioning the benchmark.
I think we would all be better served by inviting skepticism and encouraging people to develop an independent opinion of how well these benchmarks represent the tasks they care about. Concretely:
- For the failing tasks within the benchmark, are they: realistic, have a proper harness, graded properly, sufficiently disambiguated user intent? - Are models and API calls configured properly (ex: inconsistent context windows, reasoning levels). - Are the models equipped with a modern harness and minimal retrieval that is sufficiently realistic? - Does the score even make sense ex: is it just a LLM judge committee that produces a score on a 0-100 scale that isn't aligned with ground truth?
It's often the case you'll see many of these issues while looking at traces of a failing benchmark tasks, and it doesn't take long! Often the **very first trace you look at** has a serious issue.
I also don't think there is an issue with verified vs. flawed verdict. It's good to have opinions. If you look through enough of what's behind their opinions and you find that you usually agree with their verdict, then they have earned your trust, the same way that a benchmark can lose your trust.
I think this eval of benchmarks is a red pill that motivates investing in your own private evals.
- Andrew Curran
@AndrewCurran_ ·
This is extremely confusing, but an internal Claude research model also independently refuted the Exponential Time Hypothesis on October 16, with 14M output tokens and no human input. 'Anthropic researchers worked with Claude to improve the presentation, resulting in this paper.'
- Miles Brundage
@Miles_Brundage ·
Wonder what Argon’s chain of thought would say here
- Elon Musk
@elonmusk ·
🤭
- Rhys
@RhysSullivan ·
The next video I see from Daniel Lurie needs to be him on a construction site building more housing
- Elon Musk
@elonmusk ·
Sounds like Alex P🍆 knows a thing or two about being investigated 🤔
levent@__alpoge__ ·
“The core algorithm was found on October 16 by an internal Claude research model with 14M output tokens and no human input. Anthropic researchers worked with Claude to improve the presentation, resulting in this paper.”
https://github.com/anthropics/formal-math/tree/16a5ce41304fcdaea4e0be7c132e1c361dc26da5/eth
- Andrew Curran
@AndrewCurran_ ·
WSJ says Broadcom CEO Hock Tan directly intervened to get Anthropic's debt deal pushed though.
Peter Yang@petergyang ·
In the near future, instead of reading or doing anything online, people are just going to tag their agent to do it.
@bot is an early signal of this.
- dax
@thdxr ·
what is the best github alternative today for private repos
- Laurence Moroney 🇺🇸🇮🇪 🏴
@lmoroney ·
What does an open video model look like a few months after release?
MiniMax just posted a community update on H3, their open-weights video model with native stereo audio. The latest contributions include 8-step reference-conditioned video with audio, 4-step distillation, sparse attention and reference-token compression. There are also serving and workflow integrations for GPU servers, TPUs and Apple Silicon, with 11 new additions to their H3 ecosystem index.
This is the part of open weights I enjoy watching as an educator. Researchers and hobbyists improve the methods, adapt the runtime and share their workflows, and that makes experimenting cheaper for everyone learning the craft.
If you make AI films, try this: render one shot at full steps and again with a 4-step distilled version, then watch them side by side and note where the detail drops. 🎬
#AIFilmmaking #OpenSource
Link in Comments
- Theo - t3.gg
@theo ·
We now steer users away from multi-account in T3 Code directly. You can do it if you want, but I really think you should just use CLIProxy.
- Andrew Curran
@AndrewCurran_ ·
OpenAI has verified that no existing cryptosystems are compromised by today’s result. Looking at the current trendline though, Erdős, Navier-Stokes, the OpenAI 700+ GitHub dump, and today’s release were all in the last six months. It's hard to deny where this is going.
Vals AI@ValsAI ·
GPT-6.1 Terra is now #10 on the Vals Index, less than 4 points behind GPT-6.1 Sol, at half Sol's token price.
