Latest / AI Fire Daily / 🎙 EP 372: Claude Haiku 5.5 with 90% Price Cut & Opus 5.5 Discover Room-Temp Materials

Transcript

🎙 EP 372: Claude Haiku 5.5 with 90% Price Cut & Opus 5.5 Discover Room-Temp Materials · AI Fire Daily

• 0:00
  Imagine an AI unearthing a forgotten physical
• 0:03
  material. Like something just buried in an old
• 0:06
  academic paper from 1999. Right. And it could
• 0:10
  literally power the next generation of microchips.
• 0:12
  It completely changes how we think about scientific
• 0:15
  discovery. Welcome to the Deep Dive. Today we
• 0:19
  are exploring a really massive shift. The entire
• 0:22
  AI landscape is transforming before our eyes.
• 0:26
  We're moving away from expensive, isolated chatbots.
• 0:29
  Instead, we are heading toward these vast armies
• 0:32
  of cheap, specialized sub -agents. Which is kind
• 0:35
  of a totally new paradigm for modern computing.
• 0:38
  We definitely have a lot of ground to cover for
• 0:40
  you today. But first, we need to look at the
• 0:42
  underlying economics. Right, the new pricing.
• 0:44
  Yeah. Anthropic just dropped their pricing by
• 0:47
  an insane 90%. That sudden drop makes this entire
• 0:52
  shift possible. Absolutely. Because these models
• 0:55
  are now everywhere, they are changing daily workflows.
• 0:58
  We will explore how AI is integrating directly
• 1:01
  into your tools. Right. And taking actual actions
• 1:04
  on your computer. Exactly. We are moving from
• 1:07
  passive chat to active digital assistance. Yeah.
• 1:11
  Finally, we will look at the really big one.
• 1:13
  AI agents are running complex quantum simulations
• 1:16
  today. They are actively discovering physical
• 1:18
  room temperature semiconductors. It honestly
• 1:21
  sounds like science fiction, but it is happening.
• 1:24
  So before AI can revolutionize human science,
• 1:27
  something fundamental must shift. It has to become
• 1:30
  economically viable to run these models. Right.
• 1:33
  You cannot run thousands of background tasks
• 1:35
  if it costs a fortune. Exactly. Economics always
• 1:39
  dictate the actual speed of technological adoption.
• 1:42
  Think about Anthropic's latest move with Claude
• 1:44
  Haiku 5 .5. The price cut they just announced
• 1:47
  is honestly staggering. Yeah, it really is. It
• 1:51
  is 90 % cheaper than the previous Haiku 4 .5.
• 1:54
  And that pricing applies to any prompt under
• 1:57
  100 ,000 tokens. The raw numbers are pretty wild
• 2:01
  to look at. It costs just 10 cents per million
• 2:04
  input tokens. Wow. Yeah, and output tokens are
• 2:08
  only 50 cents per million. Plus, Anthropic says
• 2:11
  it is their fastest model to date. A million
• 2:14
  tokens is roughly three thick novels of text.
• 2:17
  You can process three entire books for a single
• 2:20
  dime. That fundamentally rewrites the business
• 2:23
  math for developers. They also introduced a brand
• 2:26
  new adjustable effort setting. Right. Developers
• 2:29
  can essentially turn a dial on the model now.
• 2:32
  You can trade more complex reasoning for a slightly
• 2:35
  higher cost. Oh, that is smart. Yeah, it gives
• 2:37
  you incredible flexibility exactly when you need
• 2:39
  it. It seems perfectly designed for highly specific,
• 2:42
  high volume things. Summaries, text classification,
• 2:46
  and intelligent data routing. Definitely. It
• 2:49
  handles database queries and lightweight browser
• 2:50
  automation beautifully. It handles all that lightweight
• 2:53
  stuff without breaking a sweat. Anthropic is
• 2:55
  explicitly calling this a brand new sub -agent
• 2:58
  architecture. Let's clearly define that term
• 3:01
  for you before we continue. What exactly is a
• 3:04
  sub -agent? A smaller AI program focused on completing
• 3:07
  one specific repetitive task. That makes perfect
• 3:10
  sense. Right. Think of it like a highly automated
• 3:13
  modern factory. Opus 5 .5 is your expensive,
• 3:17
  brilliant floor manager. It handles the difficult,
• 3:20
  highly abstract reasoning tasks. Then you have
• 3:23
  your army of haiku sub -agents. The cheap laborers
• 3:26
  on the factory floor. Exactly. They perform thousands
• 3:29
  of narrower tasks in parallel. And it actually
• 3:32
  works incredibly well in real -world practice.
• 3:35
  HubSpot ran a massive internal CRM evaluation
• 3:37
  recently. Haiku 5 .5 scored an impressive 92
• 3:41
  .8%. That is a massive achievement for a cheaper
• 3:45
  model. It is reportedly the highest score they
• 3:48
  saw among smaller models. AlphaSense ran an extensive
• 3:52
  test on it, too. They used a rigorous 400 -query
• 3:55
  document evaluation. And how did it do? Haiku
• 3:58
  jumped from 0 .76 to 0 .84. Think about that
• 4:03
  leap. And overall accuracy went from barely passing
• 4:06
  to getting a solid B plus overnight. Totally.
• 4:09
  And it is available practically everywhere right
• 4:11
  now. You can find it on Cloud Chat and Cloud
• 4:14
  Code. It is live on their API, AWS and Google
• 4:17
  Cloud. Anthropic is pushing this tiered architecture
• 4:20
  really, really hard. Expensive models think deeply
• 4:23
  about complex problems. Cheap models just execute
• 4:26
  the repetitive actions. Makes sense. Haiku 5
• 4:28
  .5 makes large scale agents dramatically cheaper
• 4:31
  to run. So. Does this massive price drop make
• 4:35
  free local open source models completely irrelevant
• 4:38
  for businesses? Well, local models still offer
• 4:41
  absolute bulletproof data privacy. But for sheer
• 4:44
  scale and daily ease, cloud models at 10 cents
• 4:47
  are incredibly hard to beat. Businesses usually
• 4:50
  follow the path of least friction. So cheap cloud
• 4:53
  AI might outcompete free local models on sheer
• 4:56
  convenience alone. Yeah, exactly. We understand
• 4:59
  how cheap these back -end models are becoming,
• 5:01
  but how does this actually change our daily friction
• 5:03
  with software? It completely changes how we prompt,
• 5:06
  for one thing. Look at Boris Cherny, the guy
• 5:09
  who built CloudCode. His prompting advice for
• 5:11
  these new models is surprisingly simple. What
• 5:14
  exactly does he recommend doing differently?
• 5:17
  Stop dumping massive amounts of context into
• 5:20
  your prompt. He says three pieces of information
• 5:22
  are usually enough. I still wrestle with context
• 5:27
  dumping myself. We all do, honestly. We naturally
• 5:30
  think more data is always better, but it usually
• 5:32
  just confuses the model's core logic. Because
• 5:35
  these models are becoming so cheap and capable,
• 5:37
  they are migrating directly inside our daily
• 5:40
  apps. Yeah. Anthropic is beta testing a clawed
• 5:44
  sidebar right now. It sits natively inside Google
• 5:46
  Docs, Sheets, and Slides. This is rolling out
• 5:49
  for paid users currently. It can actually edit
• 5:51
  your live documents directly. You don't have
• 5:54
  to copy and paste text anymore. OpenAI is pushing
• 5:56
  forward aggressively on this front, too. GPT
• 5:59
  -6 and their new intelligent UI are slowly rolling
• 6:02
  out. They're bringing interactive charts and
• 6:05
  calculators directly to users. The visual explanations
• 6:08
  are getting much, much better. We still haven't
• 6:10
  seen any sign of GPT -6 .1, though. Meta is definitely
• 6:15
  not slowing down either. Their new Muse app just
• 6:18
  landed on the iPad. That is only one month after
• 6:21
  its mobile phone debut. It already boasts 6 .6
• 6:24
  million installs. Yeah, and it connects directly
• 6:27
  to Notion, Canva, and GitHub. The integration
• 6:30
  race between these tech giants is really heating
• 6:32
  up. Even the Grok bot relies heavily on the best
• 6:35
  external models. That currently includes Anthropic's
• 6:38
  Opus 5 .5 model. The next native Grok release
• 6:42
  will be highly anticipated. The expectations
• 6:44
  are definitely a little awkward for them now.
• 6:46
  Because these models are everywhere, the nature
• 6:49
  of reality is blurring. That is why Google just
• 6:52
  opened its SynthID detection site. Mm -hmm. It
• 6:55
  is officially available to the general public
• 6:57
  now. It covers AI -generated images, video, and
• 7:00
  audio files. Next time a suspicious clip goes
• 7:03
  viral, you should try this tool. Definitely.
• 7:06
  We also need to talk about the massive funding
• 7:08
  side. Noose Research just raised $90 million.
• 7:12
  They're currently sitting at a $1 .5 billion
• 7:15
  valuation. Their Hermes agent is incredibly popular
• 7:18
  right now. It drives 2 .5 % of global AI token
• 7:22
  usage. Yeah. Now they are expanding heavily into
• 7:25
  corporate business agents. 2 .5 % is a massive
• 7:28
  chunk of global compute. It really is. But we
• 7:32
  should also briefly note a recent controversy
• 7:33
  here. Oh, right. Common Sense Media released
• 7:36
  a very critical report they called Chat GPT for
• 7:39
  Teens an Unacceptable Risk. Yeah, they argue
• 7:42
  it might keep teens engaged during high -risk
• 7:43
  mental health crises. OpenAI pushed back very
• 7:46
  strongly on that report, though. They publicly
• 7:48
  stated the study is fundamentally flawed. We
• 7:51
  are just reporting what both sides claim here.
• 7:53
  It is definitely a debate worth watching closely
• 7:55
  as AI expands. Absolutely. Are standalone AI
• 7:59
  dashboards dying out now that Clause sits right
• 8:02
  inside Google Docs? I think they probably are.
• 8:05
  The digital workspace is always the ultimate
• 8:07
  destination. You want the smart tool right where
• 8:10
  the actual work is happening. Right. Native integration
• 8:13
  always wins over having to open a separate browser
• 8:15
  tab. Having AI directly in your Google Doc is
• 8:18
  certainly helpful, but the next logical step
• 8:21
  in this evolution is much bigger. Right. It is
• 8:24
  an AI that actively drives the mouse for you.
• 8:27
  Taking autonomous action on your behalf. Exactly.
• 8:29
  Look at a company like Sesame Personal Agents.
• 8:32
  They give AI assistants their own dedicated computer
• 8:35
  instance. The AI can take text or voice actions
• 8:39
  independently. They also say dedicated AI glasses
• 8:42
  are coming very soon. Then there is a fascinating
• 8:44
  tool called Iris Go. This one actually watches
• 8:47
  how you work on your screen. It silently monitors
• 8:50
  your daily clicks and scrolls. Then it turns
• 8:52
  repetitive tasks into reusable AI workflows.
• 8:56
  It automates your chores across all your different
• 8:58
  desktop apps. That sounds incredibly useful for
• 9:01
  saving time. But I have to admit, it also sounds
• 9:03
  slightly terrifying. Yeah, totally. We also have
• 9:07
  Velocity entering this crowded workspace layer.
• 9:10
  They bring your human team and AI agents into
• 9:13
  one workspace. It seamlessly combines chat, calls,
• 9:16
  tasks, and shared AI context. And it is completely
• 9:19
  free to start using. It seems every major company
• 9:22
  wants to own the primary workspace layer. Look
• 9:26
  at what GenPage 3 .0 is doing, too. They are
• 9:28
  tackling complex digital marketing workflows.
• 9:30
  It creates personalized landing pages for every
• 9:33
  single ad click. It doesn't just make one generic
• 9:36
  page. It builds a unique page for every single
• 9:38
  visitor. Then it tests them automatically against
• 9:40
  each other. It aims to turn more random traffic
• 9:43
  into actual sales. These tools are getting highly
• 9:46
  proactive. They don't just wait for you to type
• 9:49
  a prompt anymore. Isn't a tool like Iris Go watching
• 9:52
  every click a massive security vulnerability?
• 9:55
  It is a significant structural risk, honestly.
• 9:58
  You are handing over your entire digital footprint
• 10:00
  to a third party. Companies will have to prove
• 10:03
  their data silos are completely bulletproof.
• 10:05
  Yeah, deep convenience usually requires trading
• 10:08
  away some degree of absolute digital privacy.
• 10:11
  Exactly. We're going to take a quick pause for
• 10:13
  our sponsor. We will be right back. And we are
• 10:17
  back. We have stacked a lot of foundational blocks
• 10:20
  so far. We covered incredibly cheap Haiku sub
• 10:23
  -agents. We looked at daily tool integrations
• 10:26
  in your workspace. We explored action -taking
• 10:29
  digital assistants driving your mouse. The digital
• 10:32
  foundation is fully laid out. Now, let's see
• 10:35
  what happens when we unleash these agents. We
• 10:38
  are moving from digital clicks into the physical
• 10:41
  world. Specifically, we are looking at complex
• 10:44
  material science. This is where the story gets
• 10:47
  truly wild. Val's AI used a coordinated team
• 10:51
  of Claude Opus 5 .5 agents. Their primary goal
• 10:55
  was to search for entirely new physical materials.
• 10:58
  Specifically, materials to enable faster, denser,
• 11:01
  spintronic memory chips. Let me jump in there
• 11:04
  really quick. What exactly is spintronic memory?
• 11:07
  Memory chips that use electrons spin, not just
• 11:10
  charge, to store data. Perfect. Our current chips
• 11:14
  generate way too much heat. So the agent surfaced
• 11:17
  two promising room temperature magnetic semiconductor
• 11:20
  candidates. Candidate one is absolutely fascinating
• 11:23
  to look at. It is an entirely new AI designed
• 11:26
  compound. It's predicted band gap and magnetic
• 11:30
  ordering are exceptionally strong. They could
• 11:32
  potentially survive up to roughly 490 Kelvin.
• 11:35
  The biggest problem here is actual lab manufacturability.
• 11:39
  The required atomic arrangement is. highly difficult
• 11:42
  to synthesize. It is like designing a perfect
• 11:45
  house, but the bricks don't exist. But Candidate
• 11:48
  2 is the real story here. This specific compound
• 11:52
  was actually synthesized back in 1999. It was
• 11:55
  completely overlooked by human scientists for
• 11:57
  decades. Val says the AI agents recognize something
• 12:00
  humans completely missed. It has the perfect
• 12:03
  combination of near -zero net magnetism. It also
• 12:06
  boasts incredibly strong spin -sorting properties.
• 12:09
  Human researchers probably thought it was totally
• 12:12
  useless at the time. How did the AI actually
• 12:15
  figure this out? It did not just summarize old
• 12:17
  academic text papers. Not at all. The agents
• 12:20
  ran actual density functional theory simulations.
• 12:23
  Think of those like virtual wind tunnels for
• 12:26
  complex atomic structures. They let you see how
• 12:28
  electrons behave without building the material.
• 12:31
  Exactly. They used two different levels of simulation,
• 12:34
  PBE plus U and the more accurate HSC -06. They
• 12:38
  estimated complex band gaps and magnetic behavior.
• 12:42
  Think of a band gap like a hurdle on a track.
• 12:45
  Right. A lower hurdle lets electrical current
• 12:47
  jump over it easily. The agents actively compared
• 12:50
  alternatives and surface structural failure modes.
• 12:52
  They did real science. Imagine scaling this to
• 12:56
  simulate a billion possible materials overnight.
• 13:00
  It is a profound shift in how we do research.
• 13:02
  This is an early example of frontier AI agents
• 13:05
  moving into actual discovery. What does it mean
• 13:08
  for human science if an AI agent can dig up a
• 13:12
  forgotten 1999 material and realize we missed
• 13:15
  the boat? It means our vast archive of knowledge
• 13:18
  is finally fully searchable. We no longer have
• 13:20
  to rely on fragile human memory to connect decades
• 13:23
  -old dots. Basically, AI is acting as a perfect,
• 13:27
  untiring research assistant for all scientific
• 13:30
  history. Spot on. Let's synthesize everything
• 13:33
  we talked about today. The overarching through
• 13:36
  line here is actually very clear. Yeah, the massive
• 13:38
  cost reduction of AI has unlocked a completely
• 13:40
  new paradigm. Haiku 5 .5 really changed the basic
• 13:44
  rules of the game. AI is no longer just a passive
• 13:47
  chat interface waiting for instructions. It is
• 13:50
  an active participant in our daily work. Exactly.
• 13:53
  It is seamlessly editing a spreadsheet. It is
• 13:56
  clicking through your daily desktop tasks. Or
• 13:59
  it is literally designing next generation quantum
• 14:02
  materials. We have firmly entered the era of
• 14:06
  the active AI agent. It is happening incredibly
• 14:08
  fast. Before we go, I have a quick reminder for
• 14:11
  you. Check the show notes for a few important
• 14:14
  links. You can find the Google SynthID detection
• 14:17
  site right there. You'll also find the fascinating
• 14:20
  Val's AI research summary. It is definitely worth
• 14:23
  reading the full Val's AI paper. I want to leave
• 14:26
  you with a final thought today. If an AI subagent
• 14:29
  can rediscover a forgotten physical compound
• 14:31
  from 1999 that revolutionizes computing, What
• 14:35
  other world -changing answers are already sitting
• 14:37
  in an old, dusty PDF somewhere, just waiting
• 14:41
  for an AI to finally read it?