Latest / AI Fire Daily / 🎙 EP 379: Google Open-Sources Harness² & Elorian Drops Glimpse Dynamic Vision

Transcript

🎙 EP 379: Google Open-Sources Harness² & Elorian Drops Glimpse Dynamic Vision · AI Fire Daily

• 0:00
  We usually assume AI is completely static. You
• 0:02
  know, you think it learns once and just stops.
• 0:06
  But what happens when agents start upgrading
• 0:08
  their own brains on the fly? Right, or when multimodal
• 0:12
  models learn to do a human -style double -take.
• 0:14
  Exactly. Welcome to the Deep Dive. Yeah, thanks
• 0:17
  for having me. Today we are unpacking a really
• 0:20
  dense stack of AI newsletter sources. We want
• 0:22
  to shortcut your way to being well -informed.
• 0:24
  We have a ton of fascinating ground to cover
• 0:26
  today. We really do. First, we'll look at Google
• 0:29
  Research's Harness Squared framework. Then we'll
• 0:32
  explore a total whirlwind of industry news. Anthropic,
• 0:35
  Disney, and some major geopolitical shifts. Right.
• 0:39
  And finally, we will unpack Elorian's fascinating
• 0:41
  new true vision model called Glimpse. It's a
• 0:44
  huge week. Okay, let's unpack this Google Research
• 0:47
  paper first. So Google Research just open sourced
• 0:50
  an amazing new framework. It's called Harness
• 0:52
  Squared. What's fascinating here is how it treats
• 0:55
  agent improvement. It actually lets AI agents
• 0:59
  improve how they work over time. And they do
• 1:02
  this without changing the underlying model weights,
• 1:04
  right? Exactly. Yeah. No way it changes at all.
• 1:07
  For those newer to this, model weights are just
• 1:09
  the core math determining how an AI makes decisions.
• 1:12
  Right. So the core intelligence stays. Exactly
• 1:15
  the same. But the way it actually applies that
• 1:18
  intelligence evolves autonomously. It rewrites
• 1:22
  its own operating software on the fly. Basically,
• 1:24
  yeah. Harness Squared can rewrite eight specific
• 1:27
  parts of an agent. Eight parts? Like what? It
• 1:31
  alters the system prompt and the safety guardrails.
• 1:34
  It also changes memory, tool guidance, and its
• 1:37
  sub -agents. Oh, wow. Yeah. It even rewrites
• 1:40
  skills, scripts, and external software plugins.
• 1:43
  This sounds incredibly complex to manage. How
• 1:46
  does it actually know what to rewrite autonomously?
• 1:49
  Well, it does this without needing any ground
• 1:51
  truth answer. Meaning a perfectly verified human
• 1:54
  response used for grading. Exactly. Instead of
• 1:57
  relying on a static human grading key, it figures
• 2:00
  out what works during the task itself. Here's
• 2:03
  where it gets really interesting. How does it
• 2:06
  actually figure that out without our help? Think
• 2:08
  about it like stacking Lego blocks of data. Okay,
• 2:12
  I like that. A normal AI just blindly follows
• 2:15
  a static set of instructions. But Harness Squared
• 2:17
  tries different Lego configurations while working.
• 2:20
  So it's testing blocks as it goes. Right. So
• 2:24
  the next time it encounters that exact task,
• 2:26
  the starting structure is already much stronger.
• 2:29
  I still wrestle with prompt drift myself. Getting
• 2:32
  an AI to stay on track is really hard. So the
• 2:36
  idea of an AI fixing its own prompts is wild.
• 2:41
  Definitely changes the whole paradigm of prompt
• 2:42
  engineering. Didn't early tests show some massive
• 2:45
  gains? Oh, truly massive. They saw a jump of
• 2:49
  11 .5 points. On which metric? That was on the
• 2:53
  work buddy bench evaluation metric. We need to
• 2:56
  contextualize that benchmark number for a second.
• 2:58
  11 .5 points is a massive leap forward. It really
• 3:02
  is. Most models struggle to gain even a single
• 3:05
  point there. Because WorkBuddyBench tests highly
• 3:08
  complex, multi -step agent workflows. Right.
• 3:11
  Getting an 11 -point jump is absolutely unheard
• 3:14
  of today. They also saw 4 .5 points on JobBench
• 3:18
  and an 11 .8 percentage point increase on Lab's
• 3:22
  partial pass rate. They tested this on Gemini
• 3:25
  3 .5 Flash, right? Yeah, and they ran experiments
• 3:28
  using the GLM -5 model, too. Just to be clear
• 3:31
  on the current availability for everyone listening.
• 3:34
  Google released this project under the Apache
• 3:36
  2 .0 open source license. Right. And it actively
• 3:39
  supports both open code and codex task executors.
• 3:43
  But we should definitely note this is just a
• 3:45
  research project. It is not an officially supported
• 3:47
  Google product yet. Yeah, that's an important
• 3:49
  distinction. So this raises a pretty big question
• 3:52
  about safety and control. Letting an AI rewrite
• 3:56
  its own guardrails feels inherently risky. If
• 3:59
  an agent removes its safety filters, things could
• 4:01
  break. Wait, I'm really stuck on that specific
• 4:04
  risk factor. If it rewrites a guardrail to be
• 4:06
  completely unsafe, how do we ensure safety when
• 4:09
  agents rewrite their own guardrails? So a persistent
• 4:12
  library acts as a strict safety buffer here.
• 4:15
  Okay. If an agent writes a bad guardrail, the
• 4:19
  system checks it. It tests the new configuration
• 4:21
  in an isolated software sandbox. Oh, I see. If
• 4:26
  the performance drops, it just reverts to the
• 4:28
  previous version. So it learns from failure without
• 4:30
  causing a system collapse. We sandbox the upgrades
• 4:33
  to keep the safety rails intact. Exactly. It
• 4:37
  gives us a secure safety net for autonomous improvement.
• 4:40
  Okay, let's transition over to the human side
• 4:43
  of things. The whole corporate industry is desperately
• 4:45
  trying to keep pace with this. Yeah, there is
• 4:47
  a massive amount of corporate reshuffling happening.
• 4:50
  We have a lot of micro -updates from the AI sector.
• 4:52
  Where should we start? Let's start with some
• 4:54
  really interesting news from Replit today. Replit
• 4:57
  says OpenAI models are now outperforming Claude
• 4:59
  models internally. That's specifically on what
• 5:02
  they call vibe -coded apps, right? Yeah. Vibe
• 5:05
  coding means software built using AI intuition
• 5:07
  instead of manual programming. Vibe coding sounds
• 5:10
  like a completely unscientific way to build software.
• 5:13
  It definitely sounds strange to traditional senior
• 5:16
  software engineers, but it reflects how developers
• 5:18
  are actually working today. They aren't writing
• 5:21
  every single line of code manually anymore? No.
• 5:23
  They are describing the vibe or core feeling
• 5:27
  of the app. They just guide the AI until the
• 5:30
  app feels right. So GPT is now the default model
• 5:33
  for Replit users? Yeah. Cloud models are still
• 5:36
  available, but OpenAI took the lead there. Okay,
• 5:39
  what about Anthropic? They just made a massive
• 5:42
  corporate commitment. Right. If we connect this
• 5:44
  to the bigger picture, it matters. Anthropic
• 5:46
  is putting 5 % of its shares into a new fund.
• 5:49
  This is a public benefit fund for nonprofits,
• 5:52
  correct? Exactly. They will add more shares as
• 5:54
  their corporate value grows. It's designed to
• 5:57
  support various nonprofit organizations globally.
• 5:59
  And it provides direct financial aid to people
• 6:01
  in need. It will also fund crucial AI job impact
• 6:05
  research moving forward. So does this pledge
• 6:08
  create a new blueprint for corporate AI responsibility?
• 6:12
  I think it absolutely sets a new standard for
• 6:14
  the industry. Tying company equity directly to
• 6:16
  public benefit is a bold move. It forces other
• 6:19
  companies to think beyond just shareholder value.
• 6:21
  Right. They have to seriously consider the societal
• 6:23
  impact of their technology. They tie financial
• 6:26
  success directly to funding human aid. Exactly.
• 6:29
  It is a very different way to structure an AI
• 6:31
  company. Meanwhile, Disney is making some massive
• 6:34
  moves in AI talent. Disney offered jobs to most
• 6:38
  of Character AI's workforce recently. Yeah, this
• 6:40
  happened after the startup officially filed for
• 6:42
  bankruptcy protection. At least 60 employees
• 6:45
  received lucrative job offers from Disney. And
• 6:47
  most of those talented employees reportedly said
• 6:49
  yes. It shows how legacy entertainment companies
• 6:52
  are absorbing these startups. Disney knows they
• 6:56
  need this technology to stay relevant tomorrow.
• 6:58
  On the geopolitical front, there is also major
• 7:00
  global news. China is proposing to restrict exports
• 7:04
  across 23 categories of AI tech. This proposal
• 7:07
  adds another complex layer to the global AI race.
• 7:11
  It really does. The restrictions would cover
• 7:13
  everything from voice synthesis to deep learning.
• 7:15
  But the proposal keeps publicly available tech
• 7:18
  completely exempt from restrictions. We are just
• 7:22
  stating the objective facts of this policy here.
• 7:24
  Right. It is just definitely something to watch
• 7:27
  as the industry evolves globally. Let's shift
• 7:29
  gears and look at some rapid capability jumps.
• 7:32
  There's a new Claude code skill that went totally
• 7:35
  viral. It turns vague machine ideas into surprisingly
• 7:38
  detailed digital prototypes. It builds these
• 7:41
  prototypes right down to the individual bolts.
• 7:44
  Yeah, it doesn't just draw a picture of a mechanical
• 7:46
  machine part. It generates the exact three -dimensional
• 7:49
  specifications for manufacturing it. That level
• 7:52
  of granular detail is pretty astonishing to see.
• 7:55
  It really is. There was also some weird fine
• 7:57
  print from Meta recently. Oh, right. If you promoted
• 8:01
  an article on Instagram, there was a catch where...
• 8:04
  Where Meta could potentially scrape your entire
• 8:06
  website. Yeah. People read the fine print and
• 8:09
  got pretty understandably upset. Because nobody
• 8:11
  wants their entire website quietly ingested for
• 8:14
  training data. Exactly. But now that specific
• 8:17
  policy has reportedly been removed entirely.
• 8:20
  Which is good. Okay, let's pivot to a fascinating
• 8:23
  research essay by Nick Camerata. He argues that
• 8:26
  one long -held AI assumption might be completely
• 8:29
  wrong. Yeah, researchers used to think models
• 8:31
  were just impenetrable black boxes. Right. But
• 8:34
  he says AI models are far easier to interpret
• 8:37
  than assumed. If we can interpret them easily,
• 8:40
  we can inherently trust them more. Exactly. We
• 8:42
  can finally understand exactly how they arrive
• 8:44
  at their conclusions. That connects perfectly
• 8:47
  to some of the new empowered software tools.
• 8:49
  We have a quick roundup of a few here. Yeah,
• 8:52
  one tool that stands out right now is called
• 8:53
  ShimCLI. That one protects your secrets and personal
• 8:57
  data from autonomous coding agents, right? Right.
• 8:59
  It acts as a strict middleware interceptor in
• 9:02
  your local environment. So if the agent tries
• 9:05
  to read a password... The tool intercepts that
• 9:07
  request and aggressively blocks the sensitive
• 9:09
  information. That's incredibly useful. What about
• 9:12
  some consumer -focused tools? Inbox Hero brings
• 9:15
  all your email accounts into one single inbox,
• 9:17
  and it uses advanced AI to sort emails and draft
• 9:20
  personalized replies. Oh, and there is that presentation
• 9:23
  tool called MyLens. Right. It gives you a fully
• 9:26
  personalized AI presentation system. It actually
• 9:28
  answers complex audience questions live during
• 9:32
  the presentation. It even creates entirely new
• 9:34
  visual slides on the spot automatically. Yeah.
• 9:37
  If someone asks a question about a specific data
• 9:39
  point, the AI just generates a dynamic chart
• 9:42
  for it instantly. That is incredibly powerful.
• 9:45
  And finally, what about developers? We have Tower
• 9:48
  18 for advanced software developers. It lets
• 9:51
  you jump straight to your Git history instantly
• 9:53
  with a single command. It makes Git workflows
• 9:56
  much faster and way more intuitive. This connects
• 9:59
  perfectly to our final major topic for today.
• 10:02
  Yeah, how AI interprets physical reality. Exactly.
• 10:05
  But first, let's take a very quick sponsor break.
• 10:08
  Sponsor. Okay, we're back. Let's unpack this
• 10:11
  fascinating new visual reasoning model from Elorian.
• 10:14
  They just released a true vision model called
• 10:17
  Glimpse. And this represents a massive shift
• 10:20
  in how artificial intelligence sees things. It
• 10:23
  is a huge departure from older multimodal AI
• 10:26
  models. Most older models only inspect an image
• 10:29
  one single time. They convert that image into
• 10:32
  tokens and then they basically just gash. Right.
• 10:35
  They reason from that one initial static digital
• 10:38
  representation. They take one quick look and
• 10:41
  then close their eyes permanently. But Glimpse
• 10:44
  gives AI the ability to actually do a double
• 10:46
  take. It keeps looking at images and videos while
• 10:49
  it actively thinks. How does it do that? Glimpse
• 10:53
  uses what Alorian calls a dynamic visual scratch
• 10:56
  space. Is that like a digital whiteboard for
• 11:00
  the AI to use? That is a perfect analogy for
• 11:03
  how it physically functions. It allows the AI
• 11:05
  to go back to the original source. So it can
• 11:09
  physically zoom into specific pixel areas of
• 11:12
  the image. Yeah, it can crop the image to remove
• 11:14
  distracting visual background noise. It can pull
• 11:17
  more video frames to check a specific sequence.
• 11:19
  Oh, wow. It can even run Python code to verify
• 11:22
  what it actually sees. It is built for a very
• 11:25
  wide variety of complex visual tasks. Static
• 11:29
  images, fast -moving video, dense scientific
• 11:31
  charts. It is specifically designed for complex
• 11:34
  physical world spatial reasoning. This active
• 11:38
  visual scratch space is an absolute game changer
• 11:41
  for accuracy. It basically gives the AI a chance
• 11:43
  to double check itself. Let's talk about the
• 11:46
  economic cost metrics for this specific model.
• 11:48
  It costs 60 cents per million input tokens to
• 11:51
  use. And it costs $3 per million output tokens
• 11:55
  generated. That pricing is actually pretty competitive
• 11:58
  for this level of deep reasoning. It is. Wait,
• 12:00
  I am curious about the economics of this double
• 12:02
  take feature. What is the compute cost tradeoff
• 12:05
  versus traditional single glance models? Well,
• 12:08
  the initial compute cost is definitely higher
• 12:10
  for the recursive double take. Because you are
• 12:13
  spending more tokens to reevaluate the same image.
• 12:15
  Exactly. But that extra upfront processing massively
• 12:18
  reduces the number of downstream errors. Right.
• 12:21
  Fixing a cascading mistake later is usually much
• 12:24
  more expensive overall. You pay more upfront
• 12:26
  compute to save on fixing mistakes. Yeah. That
• 12:29
  is the core economic tradeoff right there in
• 12:32
  practice. You are basically buying operational
• 12:34
  certainty with those extra computation cycles.
• 12:37
  And the benchmarks reflect that. In Elorian's
• 12:40
  internal tests, Glimpse performed incredibly
• 12:42
  well against competitors. It beat the GPT -6
• 12:45
  Astra model on four different distinct benchmarks.
• 12:48
  Out of 11 demanding benchmarks in total for this
• 12:51
  specific test. It won decisively on scientific
• 12:54
  chart reasoning and short -form video captioning.
• 12:57
  Though we should carefully note that Astra still
• 12:59
  led on harder tests. Right. It is important to
• 13:02
  keep the competitive landscape in perspective
• 13:04
  here. Glimpse is available right now as a limited
• 13:07
  research preview API. The truly interesting idea
• 13:10
  here is giving AI a cognitive double take. It
• 13:14
  is much closer to how humans actually process
• 13:16
  visual evidence natively. We zoom in, we rewind
• 13:20
  the tape, and we double check things. We don't
• 13:22
  just rely on a single glance to understand physical
• 13:24
  reality. We hesitate naturally. We look closer
• 13:27
  at details and we revise our thoughts. That recursive
• 13:30
  feedback loop is exactly how we build a robust
• 13:32
  mental model. Whoa. Imagine scaling this capability
• 13:37
  to a billion concurrent queries. Yeah. Imagine
• 13:40
  an AI recursively double -checking physical reality
• 13:42
  across the entire globe. The implications for
• 13:45
  autonomous robotic systems are absolutely staggering
• 13:48
  to consider. Exactly. The AI is finally mimicking
• 13:52
  that fundamental human hesitation process. It
• 13:55
  doesn't just guess blindly when it encounters
• 13:57
  visual ambiguity anymore. It stops, re -evaluates
• 14:01
  the visual data, and adjusts its internal hypothesis.
• 14:04
  That is the foundation of true artificial visual
• 14:06
  reasoning today. So what does this all mean for
• 14:10
  you listening today? It's a huge shift. Let's
• 14:12
  synthesize the core thread of our entire deep
• 14:14
  dive today. The era of static, single -pass artificial
• 14:17
  intelligence is officially over. Yeah, we are
• 14:21
  moving into a much more dynamic, recursive technological
• 14:23
  phase. We clearly saw this with the Harness Squared
• 14:26
  framework from Google. It lets an autonomous
• 14:29
  agent iterate on its own code dynamically. It
• 14:32
  improves its own internal logic without needing
• 14:34
  human intervention at all. And we also saw this
• 14:37
  recursive pattern with the new Glimpse vision
• 14:39
  model. Right, it allows an AI to actively double
• 14:42
  -check its complex visual inputs. AI is rapidly
• 14:45
  mimicking the human process of hesitation and
• 14:48
  active revision. It is learning to carefully
• 14:51
  second -guess its own initial assumptions. This
• 14:53
  is a fundamental paradigm shift in how artificial
• 14:56
  intelligence fundamentally operates. It really
• 14:59
  changes how we need to think about these complex
• 15:00
  systems. They are becoming significantly more
• 15:03
  reflective and self -correcting every single
• 15:05
  day. They are acting more like independent researchers
• 15:08
  than simple predictive text engines. They are
• 15:11
  building their own contextual understanding of
• 15:14
  the world around them. Here is a final provocative
• 15:16
  thought for you to mull over today. We know harness
• 15:19
  squared can upgrade skills without any ground
• 15:20
  truth. And we know glimpse can actively interrogate
• 15:24
  our complex visual reality. That is a fascinating
• 15:28
  scenario to carefully imagine. It would essentially
• 15:31
  create a self -improving scientific observer.
• 15:34
  What if it autonomously writes its own tools
• 15:37
  for investigating the physical world? That is
• 15:39
  a scenario that could change absolutely everything
• 15:42
  about scientific discovery. It really is. Thank
• 15:45
  you for exploring this dense stack of sources
• 15:47
  with us today. Yeah, thanks for joining us. We
• 15:50
  deeply appreciate you taking the time for this
• 15:52
  focused deep dive. We hope this analytical shortcut
• 15:55
  helps you stay well -informed and curious today.
• 15:58
  Until next time, keep questioning the complex
• 16:00
  world evolving around you.