Grok always knows what's up. nitter.app/i/grok/share/s3Cvow6OI…

I like our mission statement "understand the universe". To me, “universe” is a metaphor for “everything”, meaning the AI’s knowledge should be all-encompassing. “understand” means that we not only build AI for humans, but specifically to advance human understanding.

I love working at @xai. No pointless internal politics and virtually no meetings - just coding and shipping. I'll do another hiring push in London this month. Get your applications in now: x.ai/careers.

You can now try Grok 3 for free here on X and on grok.com. We also released a blog post with benchmarks and examples. x.ai/blog/grok-3

Creating a scalable API from scratch was a massive effort. Here are some of my favourite engineering highlights.

Immediate API roadmap: - API playground - Better docs - Auto-upgrade of rate limits - More serving capacity - Vision model (later this week) - Tokenizer API - SDKs Planned but release date tbd: prompt caching, batch inference, fine tuning, realtime info.

Many people fail to appreciate that the engineering objective isn't just to make this work. They're making this work at volume production scale, low cost, and production-level reliability. It's infinitely harder than creating a one-off prototype.

Here's a brain dump of things I worry/think about atm. AI is getting pretty good, but human-AI bandwidth is excruciatingly low. Compared to Google Search, AI queries and answers have lower entropy and higher latency, which negatively impacts effective bandwidth.

Sounds enticing? Join us! Looking specifically for experienced C++/Rust backend engineers for grok.com and the API in the bay area: job-boards.greenhouse.io/xai… The first interview only takes 15 minutes. Here are some of my retired questions for this role & interview: - Assume you want to write your own string struct. What fields would you define on the struct? - What are move semantics and why is moving a string faster than making a copy? - Explain the differences between preemptive and cooperative concurrency. - Where is the thread scheduler defined and when does it run? - Why is iterating over an array faster than iterating over a linked list even though both operations are O(n)?

A few days before we announced @xai, Elon invited the founding team and our partners to Starbase. This is also where we agreed on our mission statement. In the evening we all had an amazing dinner together and talked about engineering until late. One of my favourite memories.

We have a similar feature on our API console. The CMD+K menu is actually connected to Grok and it allows you to get stuff done more quickly. (video is sped up)

When we started the company a year ago, our objective was to achieve three advantages: A data advantage, a talent advantage, and a compute advantage. Starting today, we have all three.

Our official Grok 4 blog post x.ai/news/grok-4

The root cause of the AWS outage will be something incredibly stupid. Like a config file that has been copied unmodified for 15 years looking like: expiration_date: 2025-10-22 # Infinite future

Our biggest challenge will be to sustain this trend. Success is measured as area under the curve.

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.

I designed the @xai logo. The "i" in the first version was slightly slanted (loosely inspired by the columns of the Parthenon). Because it rendered poorly at lower resolutions I quickly updated it the current design. Tons of designers got triggered, which I found quite amusing.

It's happening

The rate at which we can improve our models and systems is more important than any one particular milestone. Grok 3 showed that we were able to go from 0 to SotA in 19 months.

Still cooking. Dinner is about to be served.

Grok 2 happened less than a year ago. Tonight we'll unveil Grok 4.

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.

You can test our new API now and we're kicking off the launch with a new model that is fast and capable. I'm very proud of the team. The entire system was written from scratch by 6 engineers in 4 months.

Me using grok.com to debug scaling issues on grok.com

I need a better profile pic. nitter.app/i/grok/share/ZEQqHfFzt…

Happy second birthday @xai

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: boards.greenhouse.io/xai/job… Full-stack: boards.greenhouse.io/xai/job…

The past couple of weeks have been the most challenging and intense since starting @xai. Building something truly useful that users love and is commercially successful is hard. I'm glad we have a lot of people here rooting for us. Some really cool tech is on the way. Stay tuned.

June 3rd was my first day at @xai. Here are some of the things I learnt over the past 6 months:

This app is too addictive. Please stop posting bangers during standard @xai office hours from 10am to 11pm. Thank you.

It's actually quite interesting to see how Google will handle this. If you had made the same statement a few years ago, your colleagues would have legitimately formed a protest with letters and walkouts until the company fired you. The protesters would use extreme hyperboles such as claiming that they fear for their physical safety, they have anxiety going to the office now, or that he can't be trusted as a scientist because science "proves" he's wrong. While society as a whole has learnt to become more accepting of different viewpoints, I find it difficult to believe that Google's debate culture has progressed at the same pace. This wasn't always the case though. Back in the day, Google had a very healthy debate culture. We had a huge number of active email lists, TGIF was a place where you could ask Larry and Sergey hard questions, and even in tough moments, things didn't get leaked. All of this changed between 2016 and 2018 where the very worst patterns of cancel culture were embraced by a small but vocal and politically active group of Googlers. One important instrument in that change was Memegen which turned from one of the happiest places on the internet into a tool for amplifying their voices and frankly disparaging the company and its leaders. All of a sudden, any means of achieving a political goal were fair: leaking TGIF, talking to Gizmodo, live tweeting leadership Q&A. Think of all the things that happened in that timeframe and that got cancelled into oblivion: Project maven, Damore, project dragonfly. The protesters won. Google embraced their viewpoints and worked hard to suppress dissent. The thing that personally hurt me was that we went from a culture of extreme trust and honesty, to one where leaders became unable to answer any questions in a public forum. All we got was corp speak. It was around that time when Larry and Sergey stopped doing TGIF. I have obviously never talked to them but at least in my mind this newly formed political activism was an important factor in their decisions to retract. Long story short: let's see if this culture still exists at Google or if it has changed.

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.

We’re quite busy building, which is why I haven’t posted in a while. Here are some @xai updates:

We released our first image generation model that was developed entirely in-house. x.ai/blog/grok-image-generat…

Refusing to learn new things is the biggest red flag you can find in an engineer. I've seen too many people claim they can't solve a problem because they "don't know JAX, C++, Spark, k8s, etc." The best engineers simply sit down and learn things until the problem is resolved.

The entire backend system is written in Rust. This includes all UI actions which are relayed from a node server to a Rust server via gRPC.

If models improve exponentially over time, then the value of trained models depreciates exponentially, too. This is interesting from a product-perspective, as you need to move very quickly in order to extract commercial value from a model before it essentially becomes worthless.

We'll release Grok-1.5 in a couple of days and will expand access to more users and more countries. x.ai/blog/grok-1.5

Grok-1 is out in the open. x.ai/blog/grok-os

I'm very sad to hear you're leaving. There are so many moments we lived through together: flying to Austin on my first day, the crazy sprint to build Grok-1, reviewing literally 15000 applications in one crazy day, uploading the grok-1 open source website and repo on a Sunday while out in wine country with Kyle & Lilly, and many more. Thank you so much for calling me in early 2023 to offer me a position on the ship. I will be forever grateful.

In other news, today was my 2nd @xai anniversary. Not a single boring day (or night; xai never sleeps™) at this company.

I'm looking for 5-10 developers who are building real products (and ship them) to collaborate on our API. Deal is you join a group chat and give regular and detailed feedback (incl. sharing prompts) and in return you get unlimited free credits. DMs are open.

Grok just got a lot faster. We have many other improvements in the pipeline that will ship soon.

You now have $25/month of free API credits until the end of the year. If you have purchased credits already, you'll get the equivalent amount in additional free credits. x.ai/blog/api

We started rolling out Grok to the first batch of users on X.

We're giving early access to our new API at today's @xai hackathon. Great to see so many people in the room.

Grok 4 will be live on the API later today

It's difficult to express just how quickly that datacenter was built. Absolutely thrilled 🚀🚀🚀

I don't have a problem with Yann - I've never even met him. His research obviously had a huge impact on my life as I started as a computer vision researcher. I unfollowed because of the rhetoric in the post below, which I want to distance myself from.

In late 2020, Tim Lillicrap and I started the computer control project at DeepMind. We wanted to create agents that interact with computers via mouse/keyboard/pixels. We (esp. @p_humphreys) made some good progress but it was early days for general agents. Time to go back.

I walked down Market Street today after seeing some posts saying the area has improved. It actually has.

Final reminder: We're hiring in PA, SF, and London: x.ai/careers

Happy birthday @xai!

Try Grok Imagine. It's the best image generation UX I have ever used. Available on Android and iOS.

At this rate, people will pay you to use their Mistral 7B APIs by next week.

Grok-2 mini is out now on X. Grok-2 (the big boi) will be rolled out very soon. x.ai/blog/grok-2

Grok 2 is keeping the GPUs hot and things are looking very promising.

Nice little update on our compute resources and Grok 2 & 3.

Done. If you're a SuperGrok Heavy subscriber, you now get $50 / month in API credits.

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 x.ai/career/

Our inference servers are designed to support multiple regions. We can run them on-premise or in the cloud (on any provider). This will be particularly helpful in the future when we want to bring low latency APIs to more regions.

The pace of innovation and improvement at Tesla is pretty incredible.

It's a fast model x.ai/news/grok-4-fast

Preview of a new shader made by @XorDev. Needs just a few tweaks before we'll ship it.

Grok 2 is out on Hugging Face.

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.

Late-stage cancel culture

The biggest threat to any democracy is the collapse of the separation of powers. In both Germany and the UK, the three branches of government, along with the media, are colluding to pursue a broken ideology that has long been rejected by the majority of the electorate.

I learnt pretty late in life that it matters very little how prestigious your uni is. I've interviewed brilliant people from schools I never knew existed. When applying to @xai, you provide a statement of exceptional work, which is a much better skills predictor than a CV imo.

We built a fully-featured usage explorer into the console app. It's similar to something you'd find at major cloud providers.

Chrome focused on user experience more than any other browser. They invented the omnibar (other browsers had dedicated search plugins and many of them), their tab support was far superior, the browser itself was significantly faster. They also created DevTools and baked it into the browser. The best thing you could get before was FireBug on Firefox, which wasn't amazing. Finally, Chrome had a much improved rendering engine, which means fewer websites looked broken due to HTML/CSS issues. Check out the launch trailer: piped.video/watch?v=RVqbY2NU… It's also interesting to see their initial ads, which showed how fast the browser is: piped.video/watch?v=nCgQDjio…

We had a pretty great team meeting today. I've never worked on a team that got this much done in such a short amount of time.

It's cookin'

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}. x.ai/blog/grok-1.5v

CVs are an incredibly weak predictor of actual performance. I primarily look at the answer to the "what exceptional work have you done?" question and try to judge if someone has a history of winning.

If you have more people than tasks, you’d naively expect the “excess” people to simply not produce any value. Most people, however, will make up tasks and in the process may actually destroy value. This applies to governments and companies.

Andree and his team deserve a lot of credit. They not only built our data center in record time, they also built A100 and H100 clusters before, which were crucial to the development of Grok-0, Grok-1, and Grok-1.5/V.

For the longest time, this was the prevailing opinion at DeepMind. However, so far, the evidence just doesn't support this hypothesis.

Grok keeps an ear on the internet's town square in real time. share.x.ai/conversation/cfd1…

My second greatest achievement in life.

Is that world-class AI infrastructure in the room with us?

You can bind an email domain to your team to simplify team management at small companies.

The work is intense, but I'm feeling really positive about our progress @xai :)

Coincidence? I think not.

We support 2-factor authentication using TouchID, security keys (e.g. Yubikey), and authenticator apps.

This was actually quite tricky to implement. The servers are replicated. We wanted to use REST instead of websockets. We couldn't use sticky routing as that would hurt graceful restarts. You need to carefully avoid race conditions between the client state and the token stream.

The people you hire should be so good they give you imposter syndrome.

sus

David Hilbert himself hypothesized about the existence of such a team. Come and be part of history.

We (@xai / @grok crew) will do a Meet & Greet opposite the main entrance of @iclr_conf at 11am local time today. Come by to chat about research, engineering, and life at xAI!

What an amazing achievement of innovation, engineering, and execution.

Our account app shows you all active sessions and allows you to log out of any devices you might not recognize. You will also get an email every time your account is accessed from a new IP address.

I'll be at @iclr_conf in Vienna later this week (Thu/Fri). If you're interested in @xai, please DM me and we'll set up some time to chat. We're looking for great engineers to join our teams in the Bay Area and London.

I'm looking for Kubernetes warriors for the @xai team in London. Anyone* who applies today will have their first interview tomorrow, technicals on Thursday, and an offer on Friday. job-boards.greenhouse.io/xai… *if you're actually good

You can now export your programs from the PromptIDE and run them locally.

I have discovered a truly marvellous live stream link, which this post is too narrow to contain

Exceptional products need exceptional models that need exceptional compute resources. @xai has built a compute advantage that will grow exponentially with the delivery of Colossus 2, unlocking next-gen models. It's why I work here and it's why you should join, too. Links below.