Latest / AI Fire Daily / 🎙️ EP 356: Gemini 3.8 Live #1 Benchmark & Shanghai AI Lab Releases Atria Dawn

Transcript

🎙️ EP 356: Gemini 3.8 Live #1 Benchmark & Shanghai AI Lab Releases Atria Dawn · AI Fire Daily

• 0:00
  Intro music. Imagine an AI that keeps chatting
• 0:03
  with you, naturally, while simultaneously launching
• 0:06
  parallel API calls to book your flights in the
• 0:08
  background. That is not a concept for next year.
• 0:12
  It just dropped today. Beat. It really, um...
• 0:16
  It just changes our entire baseline for these
• 0:19
  systems. It does. Welcome to the Deep Dive. We
• 0:21
  are very glad you're here today. We've got a
• 0:23
  lot of ground to cover. We really do. We are
• 0:25
  sorting through today's stack of AI breakthroughs
• 0:28
  to understand a massive shift happening right
• 0:30
  now. We are moving from AI that just talks to
• 0:33
  AI that actively thinks in the background. And
• 0:36
  that background thinking is definitely not cheap.
• 0:38
  It carries a heavy, heavy physical cost. We're
• 0:41
  going to look at those staggering real -world
• 0:42
  energy costs shortly. Then we'll explore some
• 0:45
  new tools interacting with digital physics. That
• 0:47
  part is wild. Yeah. And finally, we'll unpack
• 0:50
  an open -weight model capable of running entire
• 0:53
  research projects autonomously. There is a really
• 0:56
  clear through line today. Let's start with Google's
• 0:59
  newest release. It fundamentally changes the
• 1:01
  rhythm of human -AI conversation. Google just
• 1:04
  launched Gemini 3 .8 Live. And also GemEye 3
• 1:07
  .8 Live extended thinking. Yeah, and you can
• 1:10
  find this across their API, AI Studio, the app,
• 1:12
  workspace, and search. It's everywhere now. Which
• 1:15
  is a huge rollout. It is. The standard 3 .8 Live
• 1:19
  really focuses on latency. They want fast, cost
• 1:22
  -efficient voice agents. And it relies incredibly
• 1:26
  heavily on visual grounding. Let's define that
• 1:28
  really quickly. Visual grounding means connecting
• 1:31
  what the AI sees to real -world context instantly.
• 1:34
  Exactly. It roots the AI in our physical space.
• 1:37
  But the extended thinking model... That is the
• 1:40
  real paradigm shift here. How so? Well, it handles
• 1:43
  131K input and 65K output tokens. It just effortlessly
• 1:48
  accepts text, images, audio, video. And the underlying
• 1:51
  mechanism making this possible is acing tool
• 1:53
  calls. Right. Which means doing complex tasks
• 1:55
  in the background without freezing the conversation.
• 1:57
  Yeah, it completely changes the UX. It's like
• 2:00
  talking to a master chef who is actively explaining
• 2:03
  the recipe to you while simultaneously chopping
• 2:06
  vegetables and checking the oven. That is a great
• 2:08
  way to picture it. I still wrestle with prompt
• 2:10
  drift myself, so the idea of a model maintaining
• 2:12
  context while launching parallel API calls is
• 2:16
  staggering. Beat. It's a massive hurdle they've
• 2:18
  cleared. State management during real -time speech
• 2:20
  is usually incredibly fragile. And the benchmarks
• 2:23
  are solid, right? Very solid. Extended Thinking
• 2:25
  High actually hit number one on the Artificial
• 2:27
  Analysis Speech -to -Speech Quality Index today.
• 2:30
  Wow. Number one. Yeah, though it's not number
• 2:33
  one on most general benchmarks, but... The standard
• 2:36
  3 .8 live hit number two on the speech agent
• 2:38
  arena. I am looking at this continuous voice
• 2:41
  mode and I'm a bit skeptical. Does it actually
• 2:44
  think better or is it just talking faster to
• 2:46
  hide the loading time? It genuinely executes
• 2:50
  multi -step reasoning over the retrieved results.
• 2:52
  It's not just using filler audio. It actually
• 2:55
  assesses the fresh data and then it synthesizes
• 2:57
  a new conclusion before it speaks. So it trades
• 3:00
  instant answers for deeper, continuous background
• 3:02
  reasoning. Precisely. You get a much more robust
• 3:05
  response. But, you know, that constant background
• 3:07
  thinking requires an unfathomable amount of compute
• 3:11
  power. It burns so much energy. Which is shaking
• 3:13
  up the physical world. By 2035, US AI data centers
• 3:18
  could burn more natural gas than Germany and
• 3:21
  Japan combined. Let that sink in. Germany and
• 3:24
  Japan. Whoa. Imagine scaling that kind of energy
• 3:27
  burn to a billion queries a day. Two sec silence.
• 3:31
  The physical grid is really struggling to support
• 3:34
  the digital world right now. And because of these
• 3:36
  costs, monetization is accelerating fast. Right.
• 3:38
  The era of free inference is shifting. Meta is
• 3:41
  a perfect example. They are expanding paid AI,
• 3:44
  calling it Meta One, starting at $7 .99 a month.
• 3:47
  But the creator and business tiers are what you
• 3:50
  really have to watch. Those reach up to almost
• 3:52
  $500 a month. That proves these agentic systems
• 3:55
  are serious enterprise tools now. Absolutely.
• 3:58
  And that race for premium output is why OpenAI
• 4:01
  just made a massive acquisition. Glass imaging,
• 4:03
  right. Yeah. A strategic $300 million plus acquisition.
• 4:07
  They were founded by ex -Apple engineers who
• 4:10
  actually built portrait mode. Oh, wow. Portrait
• 4:13
  mode. Yeah. They use AI to bring smartphone photos
• 4:15
  to DSLR quality. OpenAI wants to own that visual
• 4:18
  input layer. Because better visual input means
• 4:21
  better reasoning. Exactly. Speaking of shifts,
• 4:24
  Google just opened Claude to all its engineers.
• 4:27
  Which is fascinating. They are keeping Anthropic
• 4:29
  close, even despite their massive investments
• 4:32
  in Gemini. It shows how pragmatic you have to
• 4:34
  be in this space. And things move fast. GPT 5
• 4:37
  .5 is officially expiring on October 14th. Yeah.
• 4:41
  Codex users have to migrate immediately. Sam
• 4:44
  Ullman actually posted recently about two AI
• 4:46
  risks worrying him, and he stressed the need
• 4:49
  to earn public trust. Public trust is the only
• 4:51
  way they avoid crushing regulation. Right. Meanwhile,
• 4:54
  Grok Imagine just rolled out precise text editing
• 4:57
  inside generated images. That is a huge quality
• 5:01
  of life update. You can edit font, size, color,
• 5:03
  alignment. It's all moving so fast. Yeah. But
• 5:05
  I have to ask. With data centers burning that
• 5:07
  much gas and meta charging 500 bucks a month,
• 5:10
  are we hitting a ceiling on what free AI can
• 5:13
  actually do for everyday users? We kind of are.
• 5:16
  There's a growing divide now. Free models will
• 5:18
  handle basic conversational stuff, but the really
• 5:21
  expensive high compute agentic models, they just
• 5:23
  cost too much to give away. Right. The truly
• 5:26
  advanced autonomous capabilities will increasingly
• 5:29
  sit behind steep paywalls. It's just the laws
• 5:31
  of thermodynamics at this point. So if companies
• 5:34
  are paying massive premiums for these advanced
• 5:36
  capabilities, what exactly are they building?
• 5:38
  They are building tools that can touch and manipulate
• 5:41
  our actual workflows and physics. Like Odyssey
• 5:44
  3. Yes. Odyssey 3 is wild. It is a physics agent.
• 5:48
  A physics agent, meaning it learns how the world
• 5:51
  works. From visual input. Right. It doesn't just
• 5:53
  read code. It visually learns physical rules
• 5:56
  and then literally controls robots, drones, and
• 5:59
  video games. It anticipates momentum and gravity.
• 6:03
  Exactly. And we are seeing this embodiment in
• 6:05
  software too. Look at Kilo Code. Kilo Code is
• 6:08
  a game changer for developers. You can run coding
• 6:10
  agents, control VS Code, and manage CLI sessions
• 6:13
  and PR reviews entirely from a smartphone. It
• 6:16
  basically untethers you from your desk. You have
• 6:18
  an autonomous developer in your pocket. There
• 6:20
  are also new workflow apps, fundamentally restructuring
• 6:23
  how we work, like SIFT. SIFT is brilliant. It
• 6:26
  turns all your messy AI chats and ideas into
• 6:29
  a living business map. Right. It connects your
• 6:32
  decisions to actual evidence and risks. It stops
• 6:35
  your AI outputs from just getting lost in a chat
• 6:37
  window. And for the builders, there is Tune.
• 6:40
  Tune gives AI builders instant off billing and
• 6:43
  analytics. You can launch paid products from
• 6:46
  day one. But I want to go back to Odyssey 3 for
• 6:48
  a second. Is a physics agent just a fancy marketing
• 6:51
  term for a drone autopilot? Not at all. Standard
• 6:53
  autopilots just follow pre -programmed, rigid
• 6:57
  mathematical flight paths. Odyssey 3 visually
• 7:00
  learns world models, so it actively adapts to
• 7:03
  actual physics in real time. It actually understands
• 7:05
  real -world physics rather than just following
• 7:07
  rigid code. Exactly. It intuits how the world
• 7:10
  works. Which is incredible. And that leads us
• 7:12
  perfectly into our final topic. But first, a
• 7:15
  quick break, mid -roll. Welcome back. We've seen
• 7:19
  AI learn physics and control code today. Yeah,
• 7:22
  the embodiment phase is here. But the final evolution
• 7:24
  is moving from an AI that helps you research
• 7:26
  to an AI that is the researcher. This is a monumental
• 7:29
  shift in computational science. The Shanghai
• 7:32
  AI Lab just released the Atria Dawn preview and
• 7:35
  it's open weight, which is a huge deal. Let's
• 7:38
  quickly define that. Open weight means the model's
• 7:41
  core architecture is freely available to download
• 7:44
  and run. Right. It's under an MIT license, full
• 7:47
  and FP8 versions. And it has a massive context
• 7:50
  window. A context window is how much data the
• 7:52
  model can hold in its memory at once. Yeah. And
• 7:55
  Adria Dawn has a staggering 256K context window.
• 8:00
  It can digest entire stacks of academic literature
• 8:02
  in seconds. But it doesn't just answer questions.
• 8:05
  You give it a research question, and it drives
• 8:07
  it to a verifiable, reproducible result. Autonomously.
• 8:11
  Right. It searches evidence, writes and executes
• 8:13
  its own code, runs experiments, inspects the
• 8:16
  results, and produces final reports. And the
• 8:18
  craziest part, it recovers from its own failures.
• 8:21
  Yeah. Its capabilities are grouped into discovery,
• 8:23
  creation, delivery, and cybersecurity. And the
• 8:25
  benchmarks back it up. They really do. It scored
• 8:27
  96 .0 on DeepSearch QA. Which is near perfect
• 8:30
  retrieval. It hit 92 .5 on BrowseComp, 77 .0
• 8:33
  on BFCLv4, and 53 .8 on AutomationBench. It actually
• 8:37
  beat or closely matched Kimi K3, GPT 5 .6 Sol,
• 8:41
  and Opus 5 on some published tasks. Which brings
• 8:44
  up a really deep question. If the model can inspect
• 8:47
  its own results and recover from its own failures,
• 8:51
  What is left for the human scientist to actually
• 8:53
  do? Well, the paper argues that humans shift
• 8:55
  to overseeing the broad direction. We validate
• 8:58
  the final output, but the AI completes the whole
• 9:01
  project. Humans shift from doing the grunt work
• 9:04
  to acting as the principal investigator. Exactly.
• 9:07
  You design the lab. The AI runs the experiments.
• 9:10
  We have covered a lot today. We really have.
• 9:12
  And there is such a clear through line here.
• 9:14
  Autonomy. Autonomy. Exactly. We started today
• 9:17
  with Gemini 3 .8 reasoning silently in the background
• 9:20
  while chatting. Then we saw the massive physical
• 9:22
  energy costs required to run these systems. Right.
• 9:25
  Which is pushing the industry toward premium
• 9:26
  subscriptions. Then we explored tools like Odyssey
• 9:30
  3, manipulating real world physics. And we ended
• 9:33
  with Atria Dawn independently executing entire
• 9:35
  research projects. The era of AI as a simple
• 9:39
  Q &A chatbot is over. The era of the autonomous
• 9:41
  digital agent is definitely here. I want to leave
• 9:44
  you with a final thought to mull over today.
• 9:46
  Beat. If an open weight model on your laptop
• 9:49
  can independently form a hypothesis, run the
• 9:52
  experiments, recover from its own errors, and
• 9:55
  publish a verifiable result. Wait, let me rephrase
• 9:57
  that. If it can do all of that on its own. Beat.
• 10:01
  whose name goes on the patent, yours, or the
• 10:04
  model's. See you next time. Outro music.