Latest / AI Fire Daily / 🎙️ EP 359: StepFun Unveils Step 5 Preview MoE & Alibaba Open-Sources Qwen-Image-2.1

Transcript

🎙️ EP 359: StepFun Unveils Step 5 Preview MoE & Alibaba Open-Sources Qwen-Image-2.1 · AI Fire Daily

• 0:00
  There is an invisible tug of war happening right
• 0:02
  now. Oh, absolutely. And it stretches from massive
• 0:05
  geopolitical dinners all the way down to the
• 0:07
  exact cost of generating a single line of code.
• 0:11
  Yeah, the whole landscape is just shifting under
• 0:13
  our feet. It really is. Welcome to the Deep Dive.
• 0:16
  I'm glad you are here with us. Today, we are
• 0:18
  tracing the shifting front lines of artificial
• 0:21
  intelligence. There is a lot of ground to cover
• 0:23
  today. We are going to look at China's StepFun
• 0:26
  Lab first. They are unlocking this new intelligence
• 0:29
  per dollar frontier. Right. The efficiency game.
• 0:32
  Exactly. Then we will map out the geopolitical
• 0:35
  chess moves happening globally. That includes
• 0:37
  a proposed US AI force. Which is a huge development.
• 0:41
  From there, we will explore hyper practical tools
• 0:44
  for your everyday life. And we will finish with
• 0:47
  a major release from Alibaba. It is a model fundamentally
• 0:50
  changing digital design. It is a massive board
• 0:53
  to look at. But all these disparate pieces are
• 0:55
  actually deeply connected. We are watching the
• 0:58
  whole ecosystem mature at the exact same time.
• 1:01
  We really are. So let's start with the foundation
• 1:04
  models themselves. The global race has shifted
• 1:07
  quite a bit recently. It is no longer just about
• 1:09
  building the biggest possible model. That era
• 1:13
  is essentially over. Now it is about building
• 1:16
  the smartest model for the cheapest price. Yes.
• 1:19
  The focus is entirely on extreme efficiency now.
• 1:22
  The brute force approach just burns too much
• 1:24
  cash. Which brings us to Stepfun. They are an
• 1:28
  AI lab based out of Shanghai. Right. Backed by
• 1:30
  Tencent and Kuai Ming Venture Partners. And they
• 1:33
  just released something called the Step5 Preview.
• 1:35
  This is a highly significant release for the
• 1:38
  industry. Step5 is a 600 billion parameter model.
• 1:42
  Let's pause there for a second. Sure. Parameters
• 1:44
  are basically the underlying settings determining
• 1:46
  how an AI thinks. So $600 billion is an absolutely
• 1:49
  massive number of connections. It is huge. But
• 1:52
  here is the catch. It uses a MoE architecture.
• 1:55
  Right. And MoE stands for mixture of experts.
• 1:58
  It is a system that only activates the specific
• 2:01
  parts it needs. Right. And that distinction is
• 2:03
  the whole ballgame. Even though it has 600 billion
• 2:06
  parameters total, only 27 billion activate per
• 2:10
  token. That keeps the actual inference cost incredibly
• 2:14
  low. I am looking at these parameter numbers
• 2:16
  and I have to admit, I am stuck. How so? If Stepfund
• 2:21
  spent the money to build 600 billion parameters.
• 2:25
  Why on earth are they only using 27 billion of
• 2:28
  them at a time? It sounds counterintuitive, I
• 2:30
  know. Doesn't that defeat the purpose of building
• 2:33
  a massive comprehensive model? Well, think of
• 2:36
  the model like a massive world -class university.
• 2:38
  The university has hundreds of different departments.
• 2:41
  It has literature, quantum physics, accounting,
• 2:44
  and biology. Now, if you walk in and ask a calculus
• 2:47
  question, you don't call an assembly of the entire
• 2:49
  faculty. You just knock on the math department's
• 2:51
  door. Exactly. You only consult the experts in
• 2:54
  that specific field. Every time a model generates
• 2:57
  a single word, doing math across 600 billion
• 3:00
  connections requires enormous compute power.
• 3:03
  It burns expensive electricity. Right. It demands
• 3:06
  high -end silicon chips. By routing your query
• 3:10
  to just the 27 billion parameters that specialize
• 3:13
  in your topic, the model runs exponentially cheaper.
• 3:16
  So it is actually about only waking up the brain
• 3:18
  cells you actually need. That is the perfect
• 3:20
  way to phrase it. It keeps the electricity bill
• 3:22
  down while keeping the answers remarkably sharp.
• 3:26
  And the pricing on this is frankly staggering.
• 3:29
  It costs about $1 per 1 million. input tokens
• 3:34
  output tokens are just two dollars and seventy
• 3:37
  cents per million it runs incredibly fast too
• 3:39
  i mean independent testing shows roughly 100
• 3:41
  output tokens per second that is well past human
• 3:44
  reading speed it also features a 1 million token
• 3:47
  context window it even has native vision support
• 3:51
  included but i have to admit something here what's
• 3:53
  that i still wrestle with keeping track of all
• 3:56
  these competing lab benchmarks myself The numbers
• 3:58
  just blur together sometimes. You are definitely
• 4:00
  not alone in that. The benchmark fatigue is very
• 4:03
  real across the industry. Every lab claims they
• 4:05
  have the shiny new high score. But StepFun is
• 4:08
  putting up serious numbers in specific, highly
• 4:11
  lucrative fields. Right. Let's look at those
• 4:13
  fields. In coding, they scored a 67 .7 on the
• 4:18
  DeepSWE benchmark. Which is fantastic. That puts
• 4:20
  them right near top -tier models like Kimi and
• 4:23
  GLM. And finance is a major standout area for
• 4:26
  them, too. They hit 66 .4 on Frontier Finance.
• 4:30
  Wow. They also scored an 83 .3 on the Duraco
• 4:33
  benchmark. Why are coding and finance always
• 4:36
  the benchmarks labs brag about? Because those
• 4:39
  are the tasks where enterprises will actually
• 4:41
  pay massive subscription fees. Ah, that makes
• 4:43
  sense. If an AI can write clean enterprise software
• 4:46
  or analyze dense corporate financial reports
• 4:49
  without hallucinating, it is worth its weight
• 4:51
  in gold. StepFund is positioning this model around
• 4:54
  something called the Pareto Frontier. It is basically
• 4:57
  the concept of getting the best possible balance
• 4:59
  of tradeoffs. Right. You want maximum intelligence
• 5:01
  paired with minimal inference costs. Right. It
• 5:04
  is like getting supercar performance on a hybrid's
• 5:06
  fuel budget. They are proving that high -level
• 5:09
  reasoning just doesn't need a supercomputer anymore.
• 5:12
  Open weights are planned for October 15, 2026.
• 5:15
  And open weights mean the core model files are
• 5:18
  free to download. But, you know, Stepfun's pursuit
• 5:22
  of cheap, massive compute isn't happening in
• 5:25
  a vacuum. Not at all. When intelligence becomes
• 5:27
  this cheap and powerful, it stops being a tech
• 5:30
  story. It becomes a national security story.
• 5:33
  The macro landscape is shifting just as fast
• 5:35
  as the technical one. Let's look at the political
• 5:38
  side of this shift. Donald Trump recently announced
• 5:40
  plans for a federal AI force. Yeah, that made
• 5:43
  waves. It is modeled directly on the Space Force
• 5:45
  framework. He also mentioned creating a future
• 5:48
  AI czar position. We still know very few details
• 5:51
  about how this will actually work, though. The
• 5:53
  budget is completely unknown right now. Right.
• 5:55
  The internal structure isn't defined yet. We
• 5:57
  don't even know where it would sit within the
• 5:59
  government hierarchy. It is a rapidly developing
• 6:01
  story, but it signals a major structural shift
• 6:05
  in how governments view compute. In related news,
• 6:08
  there is an upcoming geopolitical dinner featuring
• 6:11
  Trump and Xi Jinping. Yeah. And the guest list
• 6:14
  is heavily focused on technology leaders. Sam
• 6:16
  Altman will be there. Tim Cook is attending.
• 6:19
  Jensen Huang is also on the list. Whoa. Imagine
• 6:22
  the decisions being made at that dinner table.
• 6:24
  Yeah. The sheer concentration of computing power
• 6:26
  represented in that single room is historic.
• 6:30
  You've got the hardware, the software and the
• 6:33
  raw manufacturing supply chain sitting down with
• 6:36
  world leaders. Interestingly, Dario Amodio. from
• 6:40
  Anthropic, has not been named among the guests
• 6:42
  yet. Tracking these power dynamics is actually
• 6:45
  becoming a complex discipline itself. It really
• 6:48
  is. A researcher named Michael Adams built something
• 6:50
  called the USGovGraph. That tool is totally fascinating
• 6:54
  to click through. Most people don't truly know
• 6:56
  who holds the levers of power across government.
• 6:59
  This graph maps out the key people and the specific
• 7:02
  positions behind the scenes. It tracks how influence
• 7:05
  and policy flow in Washington. Now, on the business
• 7:08
  and security front, we also have major shakeups
• 7:11
  today. We saw a very significant security pivot
• 7:13
  this week. Right, with ZEI. Yes, ZEI open sourced
• 7:17
  their Zcode platform, but they quickly had to
• 7:20
  remove a specific feature. Yeah, there were serious
• 7:22
  privacy concerns raised by the community. The
• 7:25
  feature essentially sent encrypted project snapshots
• 7:27
  over to Alibaba Cloud. I'm going to push on that.
• 7:30
  If the snapshots were encrypted, why was the
• 7:33
  developer community so instantly outraged? Doesn't
• 7:36
  encryption make it safe by definition? It is
• 7:39
  a massive supply chain vulnerability. When a
• 7:42
  developer is coding, a snapshot captures their
• 7:45
  entire workspace environment. It captures the
• 7:47
  proprietary logic, the passwords in the code,
• 7:50
  literally everything. Even if it is locked with
• 7:53
  encryption, sending that package to a foreign
• 7:55
  cloud server means the data is out of your physical
• 7:58
  control. Oh, I see. If that encryption key ever
• 8:00
  leaks, or if quantum computing breaks current
• 8:03
  encryption standards in five years, the host
• 8:05
  server instantly has your entire source code.
• 8:08
  Wow. For major corporations, that risk is entirely
• 8:11
  unacceptable. ZDII removed the feature quickly.
• 8:13
  They stated no data was actually retained. Meanwhile...
• 8:17
  Meta is opening up their Muse platform significantly.
• 8:20
  I think they're allowing third -party connectors
• 8:21
  now. Right. They're also integrating AI payments
• 8:25
  directly through Stripe. For small businesses,
• 8:27
  that is a massive paradigm shift. It means more
• 8:30
  ways to serve customers and process transactions
• 8:32
  directly inside the AI interface itself. And
• 8:36
  the money flowing into the space remains staggering.
• 8:39
  Manius is reportedly raising $500 million right
• 8:42
  now. Yeah. This comes at a $4 billion valuation.
• 8:46
  They recently returned to independent operations,
• 8:48
  too. The startup previously topped 100 million
• 8:50
  in annual recurring revenue. Now they are actively
• 8:53
  exploring a Hong Kong IPO. So between federal
• 8:56
  AI forces, billion dollar valuations and encrypted
• 9:00
  cloud leaks, what is the core through line here?
• 9:03
  We're watching the rapid consolidation of AI
• 9:06
  as a critical infrastructure asset. It is no
• 9:08
  longer just software you buy off the shelf. Right.
• 9:11
  Compute power is now treated exactly like national
• 9:13
  security infrastructure. Nations and corporations
• 9:15
  are maneuvering aggressively. They want to control
• 9:18
  the underlying supply chains, the talent pools,
• 9:21
  and the deployment platforms. Essentially, AI
• 9:23
  is becoming a formal pillar of global statecraft.
• 9:26
  Right. And the players who control the infrastructure
• 9:28
  today will dictate the geopolitical reality of
• 9:31
  tomorrow. We are going to pause for just a brief
• 9:34
  moment for our sponsor. Insert mid -roll sponsor
• 9:37
  Reed Placeholder here. Welcome back. We just
• 9:41
  discussed massive global statecraft and trillion
• 9:44
  -dollar dinners. Now we are moving down to the
• 9:47
  micro level. The ground floor. Exactly. We are
• 9:50
  looking at how this technology is directly replacing
• 9:53
  everyday workflows for the individual. This is
• 9:55
  where the technology gets highly practical for
• 9:58
  you. Let's look at some newly empowered AI tools.
• 10:01
  First up is Aristotle. This is an AI tutor developed
• 10:04
  alongside researchers at Stanford. It is specifically
• 10:06
  designed to help kids solve problems out loud.
• 10:09
  It uses a shared interactive whiteboard space.
• 10:12
  How does an AI utilize a shared whiteboard effectively?
• 10:15
  It watches the cursor movements and the child's
• 10:18
  drawing process in real time. Oh, wow. Yeah.
• 10:21
  If a student is working on a geometry problem,
• 10:23
  the AI doesn't just read the final answer. It
• 10:25
  sees where they pause. It sees the wrong line
• 10:28
  being drawn. It then gently corrects their logic
• 10:31
  before they get completely frustrated. That is
• 10:33
  incredible context awareness. Moving from education
• 10:37
  to business, there is a tool called MySell. It
• 10:41
  runs the entire back end of a service business.
• 10:43
  It handles basically everything you hate doing.
• 10:45
  It manages the clients, tracks the deliverables,
• 10:48
  gets the approvals, and sends the invoices. Let's
• 10:50
  paint a picture of that. Imagine you run a boutique
• 10:53
  graphic design agency. Normally, you spend your
• 10:56
  Friday mornings chasing unpaid invoices. The
• 10:59
  worst part of the job. You send awkward emails.
• 11:01
  You update complex spreadsheets. Now, MyCell
• 11:04
  sees the design project as marked complete. It
• 11:07
  drafts the invoice automatically. Just like that.
• 11:10
  It emails the client. It updates your accounting
• 11:12
  software. It completely removes the administrative
• 11:15
  friction. And it starts at $299 a month. Or you
• 11:19
  can actually self -host the open source version
• 11:22
  for free. Self -hosting is a brilliant option
• 11:24
  if you have the technical chops to manage it.
• 11:26
  Next, we have Minicart. This one is pretty wild.
• 11:30
  It is. It is for e -commerce store owners. It
• 11:32
  gives you an entire AI team to run your digital
• 11:35
  storefront. Think of it like hiring a marketing
• 11:38
  manager, a logistics coordinator, and a customer
• 11:40
  service rep all at once. It handles marketing
• 11:44
  campaigns. It manages shipping logistics. It
• 11:46
  even handles angry customer replies via chat.
• 11:49
  You literally just chat with the AI to direct
• 11:51
  your entire business. Finally, we have... Answers.
• 11:54
  This tool turns deep web research into a single
• 11:58
  API call. An API is basically a bridge letting
• 12:01
  two software programs talk. Answers returns the
• 12:05
  deep research as structured JSON with all the
• 12:08
  sources cited. JSON is just a simple text format
• 12:11
  for organizing data. Right. It makes building
• 12:13
  research agents and data pipelines incredibly
• 12:15
  smooth. It's like stacking... lego blocks of
• 12:18
  data yes and that structural element is the absolute
• 12:21
  key here write down the lego block metaphor for
• 12:23
  me how does the json structure actually change
• 12:25
  the workflow well if an ai just spits out a paragraph
• 12:28
  of text a human has to read it you have to copy
• 12:30
  the name paste it into a cell copy the date paste
• 12:32
  it into another cell it is entirely manual labor
• 12:35
  Exactly. Because Answers returns JSON, the data
• 12:39
  comes back with rigid, predictable tags. Your
• 12:42
  spreadsheet software can instantly grab the name
• 12:44
  block and the date block and snap them perfectly
• 12:46
  into your database. No human translation needed.
• 12:49
  None at all. So are these tools changing how
• 12:52
  we use software or changing the nature of the
• 12:55
  work itself? They are fundamentally changing
• 12:57
  the nature of the work. The old paradigm was
• 12:59
  operating software to complete a task. You click
• 13:02
  the buttons, you drag the files across the screen.
• 13:04
  Now we are shifting to a paradigm of delegation.
• 13:07
  You interact with autonomous agents that execute
• 13:10
  multi -step workflows entirely on your behalf.
• 13:13
  We're renting digital employees instead of just
• 13:15
  buying software subscriptions. Exactly. It frees
• 13:17
  you up to just manage the overall strategy. We
• 13:20
  are going to end today by tying this individual
• 13:22
  utility back to the open source labs. We're looking
• 13:25
  at Alibaba's AI division. Yeah, they have been
• 13:28
  busy. They just released a highly practical creative
• 13:31
  tool for designers. Yes, Alibaba dropped Quinn
• 13:34
  Image 2 .1. It is a brand new 7 billion parameter
• 13:38
  open weight image model. 7 billion parameters
• 13:42
  makes it incredibly lightweight and fast to run
• 13:45
  locally. It is especially quick for processing
• 13:48
  multi -image inputs. The really unique part is
• 13:51
  what it actually accomplishes. It combines image
• 13:54
  generation and image editing into one single
• 13:57
  cohesive system. It also features native transparent
• 14:00
  image support. That means it handles RGBA images
• 14:03
  perfectly right out of the box. That native transparency
• 14:07
  is a massive technical achievement for an open
• 14:10
  model. I have to geek out about this for a second.
• 14:12
  As someone who has spent entirely too many hours
• 14:15
  in Photoshop, native RGBA support changes everything.
• 14:19
  It really understands red, green, blue, and transparency.
• 14:22
  Right. It completely eliminates the tedious workflow
• 14:25
  of masking. Everyone knows the pain of using
• 14:27
  the magic wand tool to cut out flyaway hair from
• 14:30
  a background. But this is a nightmare. For designers
• 14:32
  and e -commerce teams, this is pure magic. You
• 14:35
  can finally just generate an isolated object
• 14:37
  without that stubborn white background box attached
• 14:39
  to it. It makes layer -based edits and text replacement
• 14:41
  incredibly easy. You just drop the generated
• 14:44
  assets straight onto your canvas. It also has
• 14:46
  incredibly strong editing fidelity. It supports
• 14:49
  up to 10 different reference images at the same
• 14:53
  time. This means it preserves brand identity
• 14:55
  and specific product details far more reliably
• 14:58
  than older models. It works wonderfully for panoramas,
• 15:01
  complex infographics, and virtual clothing try
• 15:04
  -ons. Alibaba is aiming for a very specific middle
• 15:07
  ground here. They want to sit right between the
• 15:10
  massive, expensive premium models and the weaker,
• 15:13
  unreliable small ones. But let's look under the
• 15:15
  hood. Why is having generation and editing baked
• 15:19
  into the exact same model such a crucial breakthrough?
• 15:22
  Because it allows for continuous, seamless iteration
• 15:25
  without losing the original context. Usually
• 15:28
  generating an image and then editing it requires
• 15:31
  passing data between two completely separate
• 15:33
  AI systems. That transfer introduces errors.
• 15:36
  Like a bad game of telephone. Exactly. It degrades
• 15:38
  the visual quality. Having both functions in
• 15:41
  one model means the AI deeply understands the
• 15:43
  three -dimensional space of the image it just
• 15:45
  created. It fixes the image directly instead
• 15:47
  of just guessing again. Right. The generation
• 15:49
  process and the editing process share the exact
• 15:52
  same spatial awareness. Let's take a step back
• 15:55
  and look at the big picture today. The era of
• 15:57
  AI acting as just a clever parlor trick is definitively
• 16:02
  over. We are seeing a massive evolution across
• 16:05
  every single level of society. The underlying
• 16:08
  technology is just getting vastly faster and
• 16:10
  incredibly cheap. StepFun is relentlessly driving
• 16:14
  down the fundamental cost of that reasoning.
• 16:16
  Meanwhile, world leaders are now formally treating
• 16:19
  AI as a distinct branch of military and government
• 16:22
  infrastructure. Geopolitics is shifting rapidly
• 16:25
  to secure these massive compute pipelines. And
• 16:28
  on the ground floor, small businesses are transforming
• 16:30
  entirely. Everyday people are using tools like
• 16:33
  Minicart and Quen to run entire digital storefronts.
• 16:36
  It is wild. The technology has officially matured
• 16:39
  into global infrastructure. It is no longer just
• 16:41
  a novelty app on your phone. It is the new foundation
• 16:44
  of the digital economy. The sheer speed of this
• 16:47
  integration across society is the most surprising
• 16:49
  part to me. As intelligence becomes cheaper to
• 16:51
  generate than electricity, ask yourself something.
• 16:54
  Which of your most complex daily workflows is
• 16:57
  about to become a single API call? Think about
• 17:01
  that as you go through your week. Thank you for
• 17:03
  joining us on this deep dive. Keep exploring,
• 17:05
  keep questioning, and we will catch you next
• 17:06
  time.