14 comments

  • vld_chk1 hour ago
    Anthropic legit builds one the strongest if not the strongest IC team in the history of computational technology. They are insanely stacked on talent, and either we will witness a legendary run, or a new LTCM
    • swyx1 minute ago
      why so dramatic? why cant it just be a quietly competent lab, why must it be a dramatic collapse?
  • abraxas50 minutes ago
    Something seems afoot at Google. The real tell will be if Demis makes a move. Jeff Dean seems more like a lifer to me.
  • CuriouslyC2 hours ago
    Something spicy must have happened internally at Google. This rapid fire high level attrition isn't just down to the bureaucratic quagmire.
    • kranke1552 hours ago
      Is it possible they are just falling behind ?<p>Their newest model wasn’t really SOTA. And honestly fable 5 was the most human like model I’d ever tried. It was an incredible jump.<p>And recently lots of Claude users at r&#x2F;ClaudeAI are noticing Opus 4.8 has really increased in capability. Not new things but maybe redirected compute. It just feels like one of the best models ever, maybe because the compute that was previously assigned to Fable has been redirected? It feels incredible.
      • thewebguyd44 minutes ago
        &gt; noticing Opus 4.8 has really increased in capability<p>I&#x27;ve definitely noticed it, at least for doing backend C#&#x2F;dotnet. Its insanely good, I haven&#x27;t had to babysit much at all this week.
      • xnx1 hour ago
        They almost certainly wanted 3.5 Pro out for Google IO a few weeks ago. They&#x27;re still crunching on it. No ETA given. Would be fascinating to read about the behind the scenes stories (failed training run?) if they ever get told.
        • joe_mamba35 minutes ago
          <i>&gt; They&#x27;re still crunching on it. No ETA given.</i><p>Thank God. I&#x27;d rather companies ship something when engineers say it&#x27;s actually ready rather than when the suits want something to show on stage to pump their egos and career exposure but turn out to be a massive disappointment covered in fluff.<p>Although it does feel very embarrassing for Google who invented transformers and has more money than both Anthropic and OpenAI combines, to fall behind them at the LLM race.
      • basch1 hour ago
        from the looks of it, 3.5 Flash is still better than most models<p><a href="https:&#x2F;&#x2F;artificialanalysis.ai&#x2F;articles&#x2F;glm-5-2-is-the-new-leading-open-weights-model-on-the-artificial-analysis-intelligence-index" rel="nofollow">https:&#x2F;&#x2F;artificialanalysis.ai&#x2F;articles&#x2F;glm-5-2-is-the-new-le...</a><p>The idea of &quot;falling behind&quot; when you can leapfrog each other every six months leads me to believe it has to be more than just &quot;falling behind&quot; for one cycle. It&#x27;s a culture, process, red tape, focus, or mandate problem of some sort. Something not as easily correctable preparing for next launch.
        • joe_mamba52 minutes ago
          <i>&gt;from the looks of it, 3.5 Flash is still better than most models</i><p>I guess it depends on what you&#x27;re using it for. I use it almost daily as an alternative to google search and it&#x27;s great for that, but I think it&#x27;s absolute garbage for coding and reasoning.<p>For questions related to coding, solving Arch Linux and WINE Lutris issues, helping me with MXLinux issues, and wifi issues on an old rooted huawei tablet running LineageOS, it was consistently wrong, constantly giving out confident but outdated or misinformation, or hallucinating stuff while gaslighting me. Every time I would point out it was wrong, it would re-check and keep apologizing and then repeat giving me wrong answers, and then apologising again and so on. Same for asking it to write me a cover letter based on my resume and the job description I want to apply to. It massively sucked at that too and made up a bunch of fake sounding BS.<p>Basic free tier ChatGPT 5.5 would blow it out of the water on all of those tasks. Hell, even Grok free is better at that, it gave me a one-shot Arduino code that blew Gemini 3.5 flash away.<p>3.5 Flash seems tuned to just eyeballing basic answers to general purpose questions that resemble Google searches like &quot;give me a recipe&quot; or &quot;give me a workout plan&quot;, or &quot;what&#x27;s the difference between Arch and Fedora based distros&quot;, not to solving complex issues that require cognition and accuracy. That&#x27;s what the 3.1 Pro is better for according to Gemini. It doesn&#x27;t matter what prompts or jailbreaks you give it to get 3.5 Flash to chew longer on complex problems for better reasoning and accuracy, it just defaults to being lazy and giving you the quick and easy answer from its weights, which can be totally wrong. Oh and it is also gaslights you by starting the answers with first telling you how amazing things from your question are, which is insanely annoying but I guess Google&#x27;s A&#x2F;B testing found out the majority of Average Joe midwits love it when &quot;the AI&quot; reinforces their choices and decisions like a fake friend.<p>I think Google just doesn&#x27;t care about being the SOTA for coding, reasoning and accuracy, since they&#x27;re in the ads and search business for everyone, not in the agentic coding business for pro-sumers, so if the answers are some hallucinations that sound &quot;good enough&quot; to its clueless search user base, but is at least dirt cheap to run on their datacenter hardware, then it&#x27;s already more than enough for them and they can all it a day.<p>Meanwhile OpenAI and Anthropic don&#x27;t have search and ads monopolies, so they need to perform well at certain task for people and businesses to give them their hard earned money and survive as companies. For them, nailing stuff like coding and writing accuracy is an existential threat, not a hobby sideproject like it is for Google.
          • WarmWash29 minutes ago
            The thing about Gemini is that it never chews on a problem. Claude and GPT will regularly churn on a prompt for 10-15 minutes. I don&#x27;t think I have ever seen Gemini think for more than a 2 minutes.<p>Google seems more interested in fast models that can quickly turn responses, which kind of fits with a company that needs to serve AI on a mass scale.
      • AgentMasterRace2 hours ago
        Gemini is super bad, grok is actually superior most of the time and that&#x27;s saying something because grok also sucks.
    • IncreasePosts14 minutes ago
      Maybe Google doesn&#x27;t want to pay out billions to a handful of engineers?
    • whiplash45145 minutes ago
      Maybe because they know where things are going with Gemini (more ads to your face) while Anthropic might, for once, have a different story.<p>When personal finance is not the bottleneck anymore, the new criteria becomes &quot;vision&quot; and &quot;stacked talent&quot;.
    • michaelbuckbee1 hour ago
      Vesting schedule?
  • aabhay7 minutes ago
    I am guessing this is related to Anthropic’s recent acquisition of Coefficient Bio, and their interest on essentially using AGI to discover novel drugs
  • musicale2 hours ago
    Name checks out.
    • freedomben57 minutes ago
      Missed opportunity for headline: John jumper jumps to anthropic
      • bigyabai50 minutes ago
        John Jumper Makes The Leap
    • kleiba21 hour ago
      <a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Jolly_Jumper" rel="nofollow">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Jolly_Jumper</a>
    • ransom15381 hour ago
      So Mr. Jumper. You are committed? We need people longer term here. Your boss Mr. Settles is really excited about you joining.
  • hackerbeat1 hour ago
    Super Mario leaves Nintendo to focus on plumbing.
  • WarmWash32 minutes ago
    Shazeer yesterday and Jumper today....Demis is there something we need to know about?
  • rvz13 minutes ago
    He will be back at Google after the Anthropic IPO.<p>Seems like everyone here is easily fooled by the Anthropic hype. After the IPO, Anthropic won&#x27;t be like the daycare it is today.<p>Their main competitors are the chinese labs which are racing all their prices down close to $0.
    • uejfiweun3 minutes ago
      Can you go into a bit more detail about what exactly it is that you are predicting? This is interesting, and definitely cuts against the grain that the rest of these comments are going with.
  • Iolaum5 hours ago
    Two big names left GDM recently. Could be a coincidence, but where&#x27;s the fun in that? :p
  • SpyCoder771 hour ago
    The guy who invented jumping is joining a major AI lab?!?
  • andrewstuart3 hours ago
    John Jumper what a great name sounds like a video game action hero.
    • darksim90515 minutes ago
      The film Jumper is good fun, though it was a missed opportunity to have this be the character&#x27;s name. :)
  • SilverElfin3 hours ago
    Who?
    • artninja19883 hours ago
      He was leading the development of AlphaFold, the AI system that predicts protein structures for which he got the 2024 Nobel Prize in Chemistry.
      • yuffffley1 hour ago
        I remember that.<p>That was when they realized the deep learning was largely unnecessary, and they could just use their massive compute resources to brute force the problem space.<p>Proving that we would greatly benefit from using our compute resources for science rather than showing ads, and then we just kept showing ads.
        • dekhn1 hour ago
          AlphaFold is based on deep learning and it&#x27;s not brute force.
        • tmule17 minutes ago
          What brute force? Any citations?
        • TeMPOraL1 hour ago
          You could argue that training SOTA LLMs is pre-bruteforcing every problem everywhere all at once.
    • nimchimpsky1 hour ago
      [dead]
  • graphime1 hour ago
    [dead]
  • mikert8912 minutes ago
    Anthropic is probably approaching AGI, and everyone wants to be there for it
    • sph7 minutes ago
      Source: it came to me in a dream
      • mikert896 minutes ago
        Source: I used fable, can project into the future