XAI Grok 2 in August Could Beat GPT 4o and Claude 3.5

by

Elon Musk described how XAI used 24,000 Nvidia H100s to create Grok 2, an AI model that is better than OpenAI GPT 4. XAI Grok 2 will release in August.

Elon is pushing the rapid 100k data center build to catch up for GPT 5 or 5.5. XAI set up the 100,000 Nvidia H100 themselves to get it faster so they can catch up to OpenAI.

xAI contracted for 24k H100s from Oracle and Grok 2 trained on those. Grok 2 is going through finetuning and bug fixes. Probably ready to release next month.

xAI is building the 100k H100 system itself for fastest time to completion. Aiming to begin training later this month. It will be the most powerful training cluster in the world by a large margin.

The reason we decided to do the 100k H100 and next major system internally was that our fundamental competitiveness depends on being faster than any other AI company. This is the only way to catch up.

Oracle is a great company and there is another company that shows promise also involved in that OpenAI GB200 cluster, but, when our fate depends on being the fastest by far, we must have our own hands on the steering wheel, rather than be a backseat driver.

Key Features of Grok 2.0:

1. Enhanced Reasoning and Memory:
Grok 2.0 boasts an impressive 50.6% score on the MATH benchmark, showcasing its ability to solve complex math problems from grade school to high school levels. This is a significant improvement from the 48.1% score achieved by Grok 1.5.

2. Multimodal Capabilities:
With Grok 2.0, we are introducing the first multimodal model, Grok 1.5V. This model can process a wide variety of visual information, including documents, diagrams, charts, screenshots, and photographs. This opens up a whole new realm of possibilities for AI applications.

3. Improved Performance:
Grok 2.0 is built on a hardware stack never seen before, making it the most powerful, fastest AI model created in human history. This will enable Grok 2.0 to handle even more complex tasks and provide faster responses.

4. Expanded Context:
Grok 2.0 can handle contexts up to 128,000 tokens, allowing for in-depth information recall and complex problem-solving. This is a significant improvement from the 100,000 token limit of Grok 1.5.

5. New Personality Mode:
Grok 2.0 will introduce a fun mode, giving the AI a distinctive personality and making it more entertaining to interact with. This is a departure from the more serious tone of previous versions.

6 thoughts on “XAI Grok 2 in August Could Beat GPT 4o and Claude 3.5”

  1. [ asking GPT4 about the mass of SpaceX second launch tower (~26 floors, 3 floors ~400t, steel construction ~3500t) results to ~2000-5000t estimation for the steel construction (including catching arms, heavy support devices) and ~35000t for concrete foundation for launch tower, launch pad and associated infrastructure (realistic?) ]

  2. It is not clear what they are bragging about:

    – 2.5% improvement in the MATH index, with a score of 50.6%…
    For comparison, claude sonnet 3.5 gets a score of 71.1%, and gpt4o gets a score of 76.6%.

    – Increase to 128k tokens. An insignificant increase, while other models have been giving it for a long time or give much more (2 million with Gemini Pro 1.5)

    In general, it feels like they discovered America with the vision model, while the field is already moving towards complete multi-modality, of audio and video text

    And the main thing, I didn’t understand how the title is related to the content of the article “XAI Grok 2 in August Could Beat GPT 4o and Claude 3.5”?

    Sorry…

  3. Remember Netscape?
    It is very difficult to determine who will win the AI race.
    I suspect (like browsers) there will be 4-5 companies that stand out and take 90+ share of the market, with some smaller niche players excelling for specific tasks.

  4. If Grok 2 is GPT-4o level of intelligence, that would be an impressive catch up.

    It’ll likely be 2-3 months…since we all know Elon time.

    That is also when I’m expecting it to be put into all Tesla’s, that would be awesome, sit in car, say take me home, it drives you home, zero screen interaction.
    Also I think their next gen Teslabot will have Grok 2 baked in. That’s a big deal, because THAT design is the one that gets mass produced, and the only way these things will sell in high volumes, is when it can be given verbal commands, just like a human, and it completes the task.

  5. AI progress is not slowing down at all. A new model emerge and take the lead every 2 or 3 months. It’s just that some people grossly underestimated the level needed to reach AGI/ASI. We are not at 75% AGI like Thomson is saying, but more like 30%, which is insane compared to where we were 5 years ago.

    AGI prediction : 2029
    ASI prediction : 2034

    • Kurzweil predicts 2029 as well. with singularity (~ ASI) by 2045.
      Others, e.g. Kartik Gada calculates 2060+
      In the scheme of things it doesn’t really matter i.e. (range 5-30 years) over @ least 190K years since humans on earth.
      The real question (which we haven’t yet addressed) is:
      Assuming AI/bots can take over ALL human activities (not unreasonable scenario), what will humans do?
      We are a species that needs to be occupied otherwise we will go crazy or kill each other.
      As a society I haven’t seen much discussion in this area, given the imminent social change about to happen.

Comments are closed.