Latest / Today’s AI News / OpenAI’s Superapp Takes Shape Inside Codex, Anthropic Drops Opus 4.7, OpenAI Enters Life Sciences

Transcript

OpenAI’s Superapp Takes Shape Inside Codex, Anthropic Drops Opus 4.7, OpenAI Enters Life Sciences · Today’s AI News

• 0:00
  Imagine if your Swiss Army knife suddenly grew hands and started
• 0:03
  building the campfire for you. That's essentially what just
• 0:05
  happened in the AI world today, as coding assistants have
• 0:08
  officially evolved into independent operators.
• 0:13
  Welcome back to today's AI News, brought to you by Nynex
• 0:15
  Productions. Thanks for having me back.
• 0:18
  I mean it is a wild day for tech.
• 0:20
  It really, really is. So our goal for today's deep
• 0:23
  dive is to basically unpack this massive, sudden explosion of
• 0:27
  autonomous AI agents. We're going to get into the huge
• 0:30
  leaps and specialized reasoning and really look at how all this
• 0:34
  high level tech is trickling down to local laptops and daily
• 0:38
  workflows. Yeah.
• 0:39
  And we have to start with the biggest shift, right, because we
• 0:41
  are officially moving away from the whole chatbot era.
• 0:44
  Open AI essentially just merged Chat GT, their Atlas browser
• 0:48
  platform and codecs into this, well into a unified super app.
• 0:52
  Right. And what's just fascinating to
• 0:53
  me is how it's actually operating on the back end,
• 0:55
  because it's not just using hidden APIs to talk to other
• 0:58
  software. You know they added this new
• 1:00
  feature called background computer use so codecs can
• 1:03
  navigate Mac apps that don't even have an API to plug into.
• 1:07
  Exactly. And the way that works, it's
• 1:08
  just crucial to understand the AI is basically looking at a
• 1:12
  hidden continuous screenshot of your desktop.
• 1:15
  So it uses computer vision to find a submit button or a text
• 1:19
  field, and then it calculates the X&Y coordinates on your
• 1:22
  screen. Wow.
• 1:24
  Yeah, and then it literally moves your mouse to click it.
• 1:26
  It types on your keyboard. It's seeing the interface
• 1:29
  exactly like a human does. Which means literally any legacy
• 1:32
  application you have is open to AI automation now.
• 1:35
  And they paired this with a new memory capability too, right?
• 1:38
  They did, It's in preview right now.
• 1:40
  But yeah. So you could set up some complex
• 1:42
  data scraping task, close your laptop, go to lunch and codecs
• 1:45
  will just remember where it was and pick it back up days later.
• 1:48
  That is insane. Oh, and they also embedded an in
• 1:51
  app browser powered by Atlas. Plus there's inline image
• 1:54
  generation with GPT image 1.5 S developers can generate visual
• 1:59
  UI mock ups right inside their code editor without ever
• 2:02
  switching windows. Just staying right in that flow
• 2:04
  state and and the strategy is clearly working.
• 2:07
  They announced 3,000,000 weekly users today with a 70% month
• 2:11
  over month growth rate. Fable Satu, the head of codecs,
• 2:14
  he flat out said they are quote building the Super app out in
• 2:17
  the open. Right, but Open AI isn't the
• 2:19
  only one making moves today, and Thropic dropped a huge counter
• 2:22
  move this morning. Yeah, they released Cloud Opus
• 2:25
  4.7. Exactly, and instead of building
• 2:27
  a super app and thropic is just entirely focused on raw reliable
• 2:31
  reasoning. Opus 4.7 is now the top public
• 2:34
  model for agentic codec. Let's talk about the SWE Bench
• 2:37
  Pro benchmark for a second, because Opus 4.7 just hit 64.3%
• 2:42
  on it. It's a jump of over 10 points
• 2:44
  from version 4.6. It beat Open AI, GPT 5.4 and
• 2:48
  Google's Gemini 3.1 Pro. But for anyone listening who
• 2:52
  isn't a software engineer, why does the SWE bench matter so
• 2:55
  much more than the old tests? Well, older benchmarks were
• 2:58
  basically just asking the AI to write a basic Python script from
• 3:00
  scratch, right? Like a memorization test, but
• 3:03
  SWV Bench Pro makes the AI jump into a massive messy
• 3:07
  pre-existing code base from a real GitHub project.
• 3:10
  Oh gotcha. Yeah so it has to find a deeply
• 3:12
  buried bug, figure out how thousands of files interact, and
• 3:16
  fix it without breaking everything else.
• 3:18
  Hitting 64% on that is a massive indicator of complex reasoning.
• 3:22
  OK, let's unpack this because there's a bit of a catch for
• 3:25
  developers here, right with the cost.
• 3:27
  Opus 4.7 had the exact same API pricing as 4.6.
• 3:32
  Sounds like a free upgrade but people are already saying their
• 3:35
  daily API costs are spiking. Yeah.
• 3:37
  Why is it more expensive if the price is the same?
• 3:39
  It comes down to tokens. Tokens are like the syllables or
• 3:43
  building blocks the AI uses to process information.
• 3:46
  Opus 4.7 gets that high score by essentially thinking out loud
• 3:50
  before it answers. It uses new tools like an X high
• 3:53
  effort setting for clawed code and the ultra review slash
• 3:56
  command to hunt down bugs that deeper analysis generates.
• 3:59
  Way more reasoning. Tokens.
• 4:01
  So the cost per token hasn't changed, but you're just burning
• 4:03
  through a lot more of them. Which definitely drives the
• 4:06
  monthly bill up. That's a huge detail for
• 4:08
  startups. But there's another layer to
• 4:11
  this anthropic strategy. Because Opus 4.7 is just their
• 4:14
  public model, they revealed they have this gated frontier model
• 4:18
  called Mythos Preview and it scores a massive 77.8% on that
• 4:23
  same benchmark. Yeah, nearly 80%.
• 4:25
  It's wild. So if Anthropic is burning
• 4:27
  through tokens faster and hiding its best model behind closed
• 4:30
  doors, is the public actually falling behind the true AI
• 4:33
  frontier for the first time? Because with open AI having
• 4:37
  codecs run my Mac feels like I have this invisible intern
• 4:41
  working for me. Are we being shortchanged by
• 4:43
  Anthropic holding Mythos back? I wouldn't call it
• 4:45
  shortchanging, it's more about enterprise safety.
• 4:48
  I mean, an AI that can autonomously solve almost 80% of
• 4:51
  real world software bugs is incredibly powerful and
• 4:54
  potentially unpredictable. Anthropic wants to stress test
• 4:56
  Misos in highly controlled environments with exclusive
• 4:59
  partners first. They want to be the bulletproof
• 5:01
  intelligence engine for corporations, whereas Open AI
• 5:04
  wants to be the operating system for your whole digital life.
• 5:06
  It's a fascinating tug of war in the cloud.
• 5:08
  But what if you don't want an invisible intern tracking
• 5:12
  everything in a super app, and you don't want to pay Anthropics
• 5:15
  token fees? Yeah, that brings us to this
• 5:18
  simultaneous revolution happening entirely offline,
• 5:21
  right on your. Laptop.
• 5:22
  Oh yeah, the barrier to entry for running local models has
• 5:25
  just completely collapsed. Tools like a llama are taking
• 5:28
  off. Let's walk through how this
• 5:29
  works because it sounds intimidating but it's really
• 5:32
  just downloading an installer for Mac, Linux, or Windows.
• 5:35
  But my question is, how are we running these massive language
• 5:39
  models on a standard laptop with like 8 gigabytes of RAM?
• 5:43
  It's all because of a process called quantization.
• 5:45
  A language model is made of billions of parameters, numbers
• 5:48
  that define how it thinks in the cloud.
• 5:50
  Those numbers are super precise, like a decimal going out 10
• 5:52
  places. OK, quantization.
• 5:54
  Just round those numbers off. You lose a tiny bit of reasoning
• 5:57
  accuracy, but you shrink an 8 billion parameter model down
• 6:01
  into a file that's only about 3 gigabytes.
• 6:03
  Wow, so you download a lightweight model like Gemma 3
• 6:07
  through a llama and it just runs smoothly?
• 6:10
  And the ultimate test for this is turning on airplane mode,
• 6:12
  right? Just disconnect your Wi-Fi, give
• 6:15
  the AIA massive local PDF to summarize and watch it work.
• 6:19
  Florida State. No data leaves the machine.
• 6:21
  Not at all, and the community is building great workflows for it.
• 6:24
  You can open a llama's API to give it web access if you want.
• 6:27
  Or pro tip, you can point a cloud coding agent like clawed
• 6:31
  code at your local Ala model. So it does the heavy lifting for
• 6:34
  free. Here's where it gets really
• 6:36
  interesting though. Running a llama locally feels
• 6:38
  like downloading a customized brain directly into your
• 6:40
  backpack. But if we can do this highly
• 6:43
  capable stuff locally for free, doesn't that threaten the
• 6:46
  expensive API business models for Open AI and Entropic?
• 6:50
  For daily utility task, absolutely.
• 6:52
  Summarizing PDFs, basic code completion.
• 6:54
  Local models are perfect for that, but there is a ceiling.
• 6:57
  A compressed 3 GB model just doesn't have the compute power
• 7:00
  of a massive cloud engine. If you try to refactor an entire
• 7:04
  app architecture locally, the model will just hallucinate and
• 7:06
  fail. But for privacy, it's
• 7:08
  undefeated. If you're a lawyer or a doctor,
• 7:10
  you legally can't send client data to Open AI, so Obama is
• 7:15
  really the only option. Exactly.
• 7:17
  Local is for privacy, cloud is for massive reasoning.
• 7:20
  Right, but your laptop doesn't have the compute power to, you
• 7:23
  know, cure diseases. Which brings us back to what
• 7:27
  Open AI did today with a massive domain specialization.
• 7:31
  Yeah, we are officially entering the era of hyper specialized
• 7:34
  domain models. General models are trained on
• 7:36
  everything from Wikipedia to Reddit, but for profound
• 7:39
  scientific problems you need an architecture built specifically
• 7:43
  for that logic. And that's GPT Roslyn.
• 7:46
  It's the first in Open AI's new Life Sciences series.
• 7:49
  It reads scientific papers, queries databases and actively
• 7:52
  generates novel biological hypothesis.
• 7:55
  And it does this because it's trained on the structural logic
• 7:57
  of molecules. It understands biology as a
• 7:59
  distinct language. It calculates the physical
• 8:02
  constraints of RNA sequences directly.
• 8:04
  The stats on this are crazy. They did a blind test with Dino
• 8:08
  Therapeutics and Roslyn beat 95% of human scientists on
• 8:12
  biological prediction tasks. Amgen, Moderna, the Allen
• 8:16
  Institute, they're all already using it.
• 8:18
  And this is right after they released GPT 5.4 Cyber on
• 8:22
  Tuesday. Right, the dedicated network
• 8:24
  defense model. So are we moving away from the
• 8:27
  Swiss Army knife approach towards specialized scalpels for
• 8:29
  different industries? That's the perfect analogy.
• 8:31
  The Swiss Army knife is great for writing emails, but for
• 8:34
  designing RNA therapies or defending a power grid, you need
• 8:37
  a scalpel. You need a neural network
• 8:39
  optimized for those specific laws.
• 8:42
  And Speaking of defending infrastructure, there's some big
• 8:45
  news on government adoption. Reports are saying the US
• 8:47
  government is preparing to give certain agencies access to
• 8:50
  Anthropics gated Mythos AI. And to be clear, we are just
• 8:54
  impartially reporting the source material here.
• 8:56
  We aren't taking any stance on the current government blacklist
• 8:59
  or the legal battle with Anthropic, but it objectively
• 9:03
  shows that these domain expert models are being viewed as
• 9:06
  highly critical at the national level.
• 9:08
  They really are. The reasoning capabilities of
• 9:10
  Mythos are basically strategic national assets at this point.
• 9:14
  It's. Wild.
• 9:15
  So let's look at how these macro trends, agents, and specialized
• 9:17
  environments are manifesting in today's quick It's because
• 9:21
  everything is stepping out of the browser window.
• 9:23
  Autonomy is scaling so fast. First, Perplexity rolled out the
• 9:27
  Personal Computer app. It's a Mac App running agents
• 9:30
  across 20 different Frontier models, piloting its comment
• 9:33
  browser 24/7. And Windsurf 2 point O launched
• 9:37
  an agent command center. They integrated Devon directly
• 9:40
  into the IDE, so you have a head agent delegating database work,
• 9:43
  UI design and back end logic to parallel agents all at once.
• 9:47
  Then you have the 3D World models coming out of China.
• 9:50
  Tencent open sourced HY World 2.0 generating editable physics
• 9:55
  aware 3D scenes and Alibaba introduced Happy Oyster doing
• 9:58
  the same thing from multimodal inputs.
• 10:00
  And a world model is totally different from a video
• 10:03
  generator. A video generator just makes a
• 10:04
  flat 2D illusion. A world model actually simulates
• 10:07
  gravity, lighting and collision detection.
• 10:10
  You move a table in the scene and the physics update in real
• 10:12
  time. It feels like AI is leaping out
• 10:14
  of the text box. It's hijacking our mouse cursors
• 10:18
  with perplexity and windsurf and molding digital clay with
• 10:21
  Tencent and Alibaba. Yeah, generating 3D environments
• 10:24
  on the fly is going to upend interactive media and gaming.
• 10:27
  You won't need a huge team of digital artists, you'll just
• 10:30
  describe the physics you want and the model renders it.
• 10:33
  It's super futuristic, but I want to bring this back down to
• 10:36
  earth to our community AI workflows.
• 10:39
  Let's look at how everyday people are using this.
• 10:41
  This is always my favorite part, the practical application.
• 10:44
  Today's spotlight is Jerry G He's a 73 year old author from
• 10:48
  Gig Harbor, WA. He used Claude to design his
• 10:51
  book covers, format the typography and pick titles for
• 10:54
  his last two novels. But he didn't stop at just
• 10:56
  creating the books, did he? Nope.
• 10:58
  Claude is his entire marketing pipeline.
• 11:00
  He pitches local book clubs, libraries, and manages Instagram
• 11:03
  and Blue Sky. He even built his Substack
• 11:06
  following. He calls himself a Claude
• 11:08
  convert. There's such a beautiful irony
• 11:10
  there. These multibillion dollar
• 11:12
  companies are building 3D physics engines and super apps,
• 11:16
  but the most profound impact right now is empowering a 73
• 11:19
  year old creator to run a massive publishing empire single
• 11:23
  handedly. It really is, but it.
• 11:25
  Does raise a provocative question that I want to leave
• 11:27
  everyone with today. If a 73 year old author can
• 11:30
  leverage current AI to bypass entire publishing pipelines,
• 11:34
  what is going to happen when Open AI's new background
• 11:37
  computer use Super app fully integrates into these fields?
• 11:41
  Are we looking at a future where the AI doesn't just assist you,
• 11:44
  but actively manages your entire career, negotiating contracts
• 11:48
  and running your business while you sleep?
• 11:50
  Man, that is an incredible thought to end on because the
• 11:53
  Swiss Army knife didn't just grow hand today, it started
• 11:56
  building a whole new world while we weren't looking.
• 12:00
  Thank you so much for joining us for this deep dive.
• 12:02
  If you want to stay updated with this fast moving tech world,
• 12:04
  make sure you hit that subscribe button.
• 12:06
  And if you enjoyed unpacking all of this with us today, we warmly
• 12:08
  invite you to leave a five star rating for the show.
• 12:10
  We'll see you back here for tomorrow's episode.