6 comments

  • horacemorace19 minutes ago
    I was just using this one today for some coding tinkering and it was super slow. But the way it solves problems with a light touch while still driving toward sensible architecture is refreshing after claude’s bulldozer attitude. It was able to integrate repeated creeping scope nicely with what it had already done. Very different feel than sonnet/opus, and even gemini 3.8.
  • tristanj10 minutes ago
    It&#x27;s pretty obvious that Gemini 4 Carbon is not token efficient. It also has relatively low tokens-per-second compared to other models. I can think of several reasons why this is the case:<p>1. Google has huge amounts of cash and compute. Their goal was to train a model that that has the highest quality &amp; frontier performance, not create a model that&#x27;s efficient to run and serve. They have other models for that, like Hydrogen.<p>2. Google created Carbon without caring for inference costs. DeepMind is subsidized by the rest of Google&#x27;s empire. Selling inference is an insignificant part of Google&#x27;s revenue. But for Anthropic and OpenAI, selling inference is their entire business model. Anthropic and OpenAI have enormous incentive to optimize model serving and inference costs, far more so than Google does.
  • lp926 minutes ago
    &gt;Carbon defended against cases the ticket&#x27;s facts had already excluded, hardening the code against problems nobody had reported.<p>That&#x27;s pretty interesting. Seems like it&#x27;s over analyzing. Not the worst outcome though to have additional hardening.<p>With that said, I found Argon to be better for most coding tasks and Carbon for architecture&#x2F;deep RCA analysis of complex issues.
  • jrflo12 minutes ago
    Is this due to Google wanting to catch up on benchmarks at all costs, and trains their models to be overly meticulous to do so?
  • SubiculumCode32 minutes ago
    That isn&#x27;t a good look, and either this is unmeasured uncontrolled behavior, or intentional under some longer term plan...but the latter possibility seems improbable.
    • ahsillyme17 minutes ago
      Uhm, I&#x27;m not a programmer but why isn&#x27;t this a good look? I would want my engineers to fix problems that they notice in passing. (Edit: the &quot;unnecessary work&quot; angle is what irked me in the article)
  • user_783234 minutes ago
    I am increasingly more convinced every day that every LLM company is just making their latest model to be the best at (coding) benchmarks, at the cost of pretty much everything else.<p>I use LLMs for non code purposes (probably not very common on HN) and Sonnet 3.5&#x2F;3.6 still ranks as probably one of the best models I&#x27;ve used.<p>Can today&#x27;s models put out tech demos and call more tools that make it way more impressive? Sure. But it&#x27;s really hard to see any actual <i>significant</i> gains over the last few months, they&#x27;re so mild (again, non coding).
    • incognition21 minutes ago
      Oddly flash 3.8 for me is good. they really pushed carbon to some other extreme