These are some of the UI features in Grok. First, it allows you to multi-task. You can run several concurrent conversations and switch between them as they progress.
You can now submit feedback to help us improve Grok. Feedback can range from a simple "this conversation is good/bad"-rating all the way to an entirely rewritten conversation.
Do you want to work for
@xai
in London? Now you can. We're looking for software engineers. Apply if you want to get stuff done, work with smart people, and get grilled in one of my coding interviews.
Backend & data:
Full-stack:
xAI is a great place to work at. Our culture is engineering-driven and fast-paced with a long-term strategic roadmap to develop safe and strong AI. If you want to learn more, I'm happy to jump on a Google Meet call at short notice. My DMs are open.
I'm super excited to work with the great people who are joining
@xai
today and over the next couple of weeks. Some of the smartest engineers I ever got to interview.
PS: We're still hiring
One unpopular opinion I hold is that scripting languages (Python, JS, PHP etc) are great for learning programming but insufficient for learning how computers and operating systems actually work. Becoming proficient in C or C++ can be eye-opening and exciting.
If we don't keep raising awareness of low-level programming skills, the skillset will disppear forever under a sea of SDKs, binary blobs and proprietary chips.
C programming matters, assembly language matters.
PromptIDE highlights: The `user_input()` function asks the user to enter a string. You can use that to write a chatbot in four lines of Python. While executing the prompt, you see the tokenization, sampling probabilities, alternative tokens, and aggregated attention masks.
We may finally be at the point where "super-human performance" actually means "better than most humans" in real-world applications. Can't wait for nearly flawless speech recognition, image understanding, and reasoning/deduction.
Some early results of our first vision model. It'll be integrated into the Grok chat in the medium term. A few other features will ship before that (likely very soon).
Props to {
@tingchenai
,
@gabriel_ilharco
}.
You can branch conversations to better explore Grok's responses. The response tree (wait until the end of the video) allows you to switch back and forth between the individual branches. There are also /commands to reduce clicking.
Skills, expertise, and work ethics follow a Pareto distribution with α >> 1 meaning that - in a professional context - some people are just exponentially better than others and they are rare. You want to work with and hire people who are exceptional. No compromises.
We're looking for a designer/frontend engineer to join the London team. If you're great at UI design and you can implement your designs in React, please apply so we can pixel-peep together.
Companies are groups of people. You cannot talk to a building or a machine. If something is moving slowly, there is a person who can accelerate it. If a policy doesn't make sense, there is someone who can fix it. Want to understand a company? Find out who does what.
You can only move fast if you can iterate quickly. For research, this also means that your infrastructure must support fast iteration. Launching experiments shouldn't take more than 1 minute, not every piece of code needs to be tested & reviewed, compute should be plentiful.
Innovation is paramount and you're never done innovating. It doesn't matter what you or your company achieved yesterday. The value of past rewards is 0. All that matters is how much more you will achieve tomorrow.
You can open Grok's response in a markdown editor, save it and then continue the conversation. This works in tandem with branching and the branch tree I showed above.
Thanks to everyone who applied via our website. I met a lot of people this week and really appreciate how many talented engineers and researchers sent in their CVs 🙏
Metaphorically speaking, progress made by a group of people is described by the vector sum of its members' actions. The best way to make sure a group moves quickly in one direction is to align incentives.
I refactored error management in the API. Error messages should be more accurate and less scary now. I also increased the default token limit to 100,000 tks/h. Thanks for beta testing everyone 🫡.
When interviewing, I usually try to assess a candidate's knowledge ceiling instead of checking if they pass a knowledge floor. On a few occasions, I've met people who are simply so good that I ran out of questions.
@tapfornerd
It's honestly too early to say how we'll scale this in the future. Grok launched just 10 days ago. Right now, our main focus is on making Grok as good as we can. Give us some time :)
@imPenny2x
It's very early days for the API. But if you want to play with the model programmatically, you can use our IDE. Here are a few examples to get you started:
- Chat:
- Concurrency:
- Batch process a CSV file:
You can upload small files and use them in your prompts via the `read_file()` function from the SDK. The video shows how to run an MMLU evaluation inside the IDE. The model here is our smaller Grok-0.
Strong foundation models relate to domain-specific expert models like CPUs relate to ASICs. One is general and allows its users to choose the application. The other performs well in a narrow domain chosen by the developer.
"Solve A, then use A to solve B" rarely works as it introduces a discontinuity in the solution space. You risk getting stuck iteratively working towards A and never attending to B. In my experience it's better to tackle B directly in a way that allows taking incremental steps.
@BenjaminDEKR
@tingchenai
@gabriel_ilharco
Haha, yes, we'll have an API. Developers will also get access to the IDE and our current SDK with a small token budget included with X Premium+ while we work on the fully featured API.
AI has one aunt, three fathers, three mothers, and seven godfathers. They're all talking about parenting while it hangs out with the rest of us chronically-online folk on Twitter.
@bayeslord
@grok
The Grok frontend is written in React using NextJS. The PromptIDE is written in Angular. They both share a gRPC API backend written in Rust.
I'm always happy when I meet others who worked at
@RWTH
's Computer Vision group. I learnt a lot during my time there as they allowed me to pursue my own research ideas rather than just acting as an assistant to more senior folk.
Over the past 10 years we somehow went from developers wanting to write everything from scratch to developers wanting to use SaaS-solutions for everything.
Small anecdote: The ResNet paper [1] came out while I wrote my Master's thesis. The results were impressive and the simplicity of the technique allowed everyone to validate its effectiveness within an hour or two of training. I started working on ResNets for semantic segmentation…
@DrJimFan
Congratulations to him and MIT indeed. When I focused on computer vision during my Bachelor's and Master's, his modern contributions to the field were as important to me as the early works of
@ylecun
et al.
@imPenny2x
Thanks for the kind words and the useful feedback. Lots of improvements are in the works or have already shipped silently. We'll keep pushing.
@JoshuaSteinman
@elonmusk
More seriously [2/2].
I read the discussions last night and wanted to tweet something funny about it but then didn't think anyone would care.
The i is slanted to better distinguish the three letters. I also like how it gives the logo a little bit of perspective. Though I'm not…