Latest / AI Fire Daily / 🎙️️ EP 366: OpenAI Unveils Dots & GPT-6.1 Sol & Liquid AI Launches d1 Decision Model

Transcript

🎙️️ EP 366: OpenAI Unveils Dots & GPT-6.1 Sol & Liquid AI Launches d1 Decision Model · AI Fire Daily

• 0:00
  Think about the sheer petty genius of this move.
• 0:04
  OpenAI launches their biggest new product of
• 0:06
  the year. Right. The absolute biggest launch
• 0:08
  they have. And Elon Musk's XAI quietly buys Dots
• 0:12
  .com. I mean, they did it purely to troll them.
• 0:16
  Purely to troll their biggest rival. It just
• 0:18
  redirects straight to Grokbot. It is genuinely
• 0:21
  diabolical. Welcome to today's deep dive. We
• 0:24
  are unpacking a massive shift in the AI landscape
• 0:27
  today. Yeah, we really are. It is moving so fast.
• 0:30
  We start with OpenAI's Dev Day 2026. Always -on
• 0:34
  agents took the center stage. Then we navigate
• 0:37
  a wild $1 .4 trillion ecosystem explosion. The
• 0:41
  capital flowing right now is just unbelievable.
• 0:43
  It really is. And finally, we finished with LiquidAI's
• 0:47
  new D1 model. It ignores chat entirely. It focuses
• 0:50
  purely on split -second software decisions. Let's
• 0:53
  dive right in. So OpenAI is pushing super aggressively
• 0:57
  into always -on agents. I mean, they want to
• 0:59
  be the infrastructure. The absolute base layer
• 1:02
  of everything. They announced DOTS. These are
• 1:05
  basically persistent agents living inside ChatGPT.
• 1:08
  They work across all your connected apps seamlessly.
• 1:12
  They use cloud computers for ongoing tasks. It
• 1:15
  is wild. It acts like an operating system for
• 1:18
  digital workers. That is exactly it. And that
• 1:21
  brings us to the agent's API. It is a fully managed
• 1:24
  runtime for building these digital workers. It
• 1:27
  handles persistent sessions now. Right. So the
• 1:30
  AI remembers things over long periods. Plus,
• 1:32
  the orchestration capabilities are deep. They
• 1:36
  can use custom tools and browsers directly. And
• 1:39
  they connect using MCP. Which is simply a standard
• 1:42
  way for AI to securely connect to outside data.
• 1:45
  Exactly. Securely accessing outside data is huge
• 1:48
  for enterprise. But the subagents feature is
• 1:50
  the absolute craziest part. One primary agent
• 1:53
  can spin up specialized subagents instantly.
• 1:56
  Just working in parallel. Yeah, solving complex
• 1:59
  multi -step problems. Imagine scaling to a billion
• 2:02
  queries. The computational scale required for
• 2:05
  that is truly staggering. To run this, they introduced
• 2:09
  GPT -6 .1 Sol. which supposedly hits GPT -6 Astra
• 2:14
  -level capabilities across the board. Specifically
• 2:17
  for complex coding and intense professional work.
• 2:20
  But here is the real kicker. It costs just one
• 2:22
  -fifth the token price. And a token is simply
• 2:25
  a tiny chunk of words and AI processes. Right,
• 2:28
  so dropping that price changes the startup math
• 2:30
  completely. You can finally afford to run background
• 2:33
  systems. They also announced an ultra -fast premium
• 2:37
  speed tier. Because always -on agents cannot
• 2:40
  just hang for 10 seconds while you wait. They
• 2:42
  need real -time reactions. There is also the
• 2:45
  new Pro 500 plan. It costs $500 a month. Massive
• 2:49
  usage limits. Much faster compute access. They
• 2:53
  are also shifting how we build software with
• 2:56
  Codex in the cloud. Yeah, running reusable cloud
• 2:59
  dev environments straight from a mobile phone.
• 3:01
  And chat GPT space, a shared virtual environment
• 3:04
  for humans and agents. Real -time creative collaboration
• 3:07
  with AI. It is a huge paradigm shift. But going
• 3:11
  back to that domain drama, OpenAI dropped the
• 3:13
  name DOTS without actually owning the domain.
• 3:17
  Such a massive, I mean, embarrassing oversight
• 3:19
  by their branding team. Which XAI capitalized
• 3:22
  on. So looking at that Pro 500 plan, is OpenAI
• 3:27
  pricing out indie developers? Well, it definitely
• 3:30
  raises the barrier, but they are explicitly targeting
• 3:33
  massive enterprise deployments, not individual
• 3:35
  weekend tinkerers. So it is purely for enterprise
• 3:38
  scale, not solo tinkerers. And that base layer
• 3:41
  infrastructure is triggering a total gold rush.
• 3:44
  The financial numbers are staggering. OpenAI
• 3:47
  is discussing a $30 billion funding round. Right
• 3:51
  ahead of their repeatedly delayed IPO. Valuing
• 3:53
  the company at roughly $1 .4 trillion. Investors
• 3:57
  are just eagerly piling in right now. But I have
• 4:00
  to push back here. Is a $1 .4 trillion valuation
• 4:03
  purely speculative hype? I mean, it sounds like
• 4:07
  it, but look at the actual agent boom. Instinct
• 4:11
  just raised a billion dollars. At a $10 billion
• 4:14
  valuation. Yeah. And they were valued at $2 .5
• 4:17
  billion just a single month ago. Instinct is
• 4:20
  the viral agent that autonomously books your
• 4:22
  trips and pays bills. Exactly. It navigates the
• 4:24
  web exactly like a human. That is real measurable
• 4:27
  market traction. But not everyone wants pure
• 4:30
  autonomy yet. Contrast that with Foe. Right.
• 4:32
  Launched by X Deep Mind lead Shivani. It is a
• 4:36
  hybrid personal agent. Yeah, it seamlessly tags
• 4:39
  in human assistance whenever the AI gets stuck.
• 4:42
  That might build consumer trust much faster.
• 4:44
  Oh, for sure. Big tech is expanding rapidly into
• 4:47
  this space, too. Meta is bringing its Muse AI
• 4:50
  to small businesses. Integrating with Shopify,
• 4:52
  Slack, and Dropbox. Stripe, Notion, Canva. It
• 4:56
  basically handles ops, sales, and marketing.
• 4:58
  Model selection is getting incredibly complex,
• 5:01
  though. Anthropic just published a guide. Yeah,
• 5:03
  on exactly when to use Sonnet 5 .5 versus Opus.
• 5:07
  They also basically gave GLM 5 .3 free PR. Right,
• 5:11
  benchmarking it on ExploitBench. GLM 5 .3 had
• 5:15
  50 successful runs. Against Mythos Preview's
• 5:18
  56. And Anthropic openly noted their own safeguards
• 5:22
  were much easier to bypass. Security is becoming
• 5:24
  an existential concern today. Which leads to
• 5:27
  a serious industry warning. Yeah, top leaders
• 5:29
  from Anthropic and OpenAI are sounding the alarms
• 5:32
  loudly. Warning of an impending intelligence
• 5:35
  explosion. Driven entirely by self -improving
• 5:38
  systems. So are we nearing an actual intelligence
• 5:41
  explosion or is this just industry fear -mongering?
• 5:44
  Self -improving AI makes exponential capability
• 5:47
  growth and unavoidable mathematical certainty.
• 5:50
  So the warnings are very real. Runaway capability
• 5:53
  gains are a true mathematical certainty now.
• 5:55
  Sponsor. We just discussed massive God models
• 5:59
  and intelligence explosions. Which are equal
• 6:01
  parts amazing. And terrifying. But for developers,
• 6:05
  using massive models to make tiny software choices
• 6:07
  is highly inefficient. And way too slow. That
• 6:10
  brings us to Liquid AI's D1. Their very first
• 6:12
  decision model. It is designed purely for fast,
• 6:15
  structured choices inside software. Right. It
• 6:17
  is not for long -form chatbot responses at all.
• 6:20
  It outperforms Jev on Hugging Face's decision
• 6:22
  index. And Jev was tested on over 150 ,000 scorable
• 6:26
  requests. Covering decision selection, reading,
• 6:28
  and logical inference. But there is a... Crucial
• 6:31
  caveat here. Beating Jev does not mean D1 is
• 6:34
  smarter than Frontier LLMs. It just dominates
• 6:38
  the specific bottleneck of quick, reliable micro
• 6:40
  decisions. Yeah, it is built specifically for
• 6:43
  the backend. It acts as a hyper -efficient traffic
• 6:45
  cop, not a philosopher. That is exactly it. It
• 6:49
  outputs things software can consume directly.
• 6:51
  Ranking options and routing network requests
• 6:53
  instantly. It is strong on multilingual tasks
• 6:56
  and longer contexts, too. And it is available
• 6:59
  via the Liquid API today. Open router support
• 7:02
  is coming soon. Security is a massive deal here,
• 7:05
  too. It is highly resistant to prompt injection.
• 7:07
  Which is simply malicious text designed to trick
• 7:10
  AI into executing bad actions. Right. And that
• 7:13
  innate resistance is critical when software agents
• 7:16
  make decisions on untrusted inputs. I still wrestle
• 7:19
  with prompt drift myself. Keeping models on track
• 7:22
  is difficult. Oh, it is incredibly hard. Untrusted
• 7:25
  inputs can completely derail an autonomous agent.
• 7:28
  Will specialized models eventually replace general
• 7:30
  models for all back -end tasks? Absolutely, because
• 7:33
  they are significantly faster and vastly cheaper
• 7:36
  for routing basic system logic. Specialized models
• 7:40
  win the back -end on pure efficiency. So when
• 7:42
  you step back, the entire AI landscape is bifurcating
• 7:46
  beautifully. On one end, you have massive, persistent,
• 7:50
  always -on agents. Like OpenAI's DOTS and Instinct.
• 7:54
  Fueled by absolute trillions in capital. On the
• 7:57
  other hand, hyper -specialized, invisible infrastructural
• 8:00
  models. Like Liquid's D1, quietly directing the
• 8:03
  digital traffic. AI is transitioning from a chat
• 8:06
  box to an invisible utility. It is just running
• 8:10
  the background of your life silently. Which leads
• 8:13
  you with a thought to mull over. If your personal,
• 8:16
  always -on agent from OpenAI has to negotiate
• 8:19
  with a company's back -end AI, powered by Liquid's
• 8:21
  D1, who is actually making the decisions in your
• 8:25
  life? Thank you for taking this deep dive with
• 8:29
  us today. See you next time.