Latest / Tech Talks With Kinsoft / Last Week in Tech – Anthropic vs Alibaba, OpenAI's Own Chip, and an Emergency Cisco Patch

Transcript

Last Week in Tech – Anthropic vs Alibaba, OpenAI's Own Chip, and an Emergency Cisco Patch · Tech Talks With Kinsoft

• 0:00
  Imagine finding out that the biggest threat to
• 0:03
  a multi -billion dollar AI isn't some highly
• 0:06
  sophisticated nation state super virus. Right.
• 0:09
  But instead it's just a quiet army of like 25
• 0:12
  ,000 bots asking it to write cunt for pennies.
• 0:14
  Yeah, it's wild. Or that the key to breaking
• 0:19
  a trillion dollar hardware monopoly is essentially
• 0:21
  just a software universal translator. Welcome
• 0:25
  to this custom deep dive. Today our mission for
• 0:28
  you listening is to really look underneath the
• 0:30
  hood of the AI applications you use every single
• 0:33
  day and well... examine this massive structural
• 0:36
  shift that's happening right now. Yeah. We're
• 0:38
  working from a really fascinating, incredibly
• 0:40
  dense dispatch today from a technology update.
• 0:43
  Right. Tech Talks with Kinsoft. Exactly. From
• 0:45
  June 29th, 2026. Yeah. And I mean, this source
• 0:49
  covers just one week in the tech industry. Yeah.
• 0:52
  But it acts like this perfect time capsule. So,
• 0:54
  okay, let's unpack this. Because reading through
• 0:56
  this dispatch, you realize really quickly that
• 0:57
  the, you know, the experimental novelty phase
• 1:01
  of AI, it's completely over. Oh, completely.
• 1:03
  It's done. We are looking at high stakes battlegrounds
• 1:06
  over intellectual property, massive hardware
• 1:10
  monopolies, and basically the dawn of automated
• 1:14
  cyber warfare. It is a really dense week of news.
• 1:18
  And I think the through line connecting all these
• 1:20
  seemingly separate events is really just one
• 1:23
  word, which is infrastructure. Infrastructure.
• 1:25
  Right. Yeah. Everything we're going to examine
• 1:27
  today really demonstrates how the very foundation
• 1:31
  of artificial intelligence is just being violently
• 1:33
  shaken up and rebuilt. Yeah, rebuilt from the
• 1:36
  ground up. Well, we are seeing a complete reconstruction
• 1:38
  happening on every single level. You know, how
• 1:41
  the underlying training data is safeguarded,
• 1:43
  how the physical silicon itself is manufactured,
• 1:46
  and how the entire security architecture of our
• 1:49
  networks is fundamentally changing. Well, I want
• 1:52
  to start with the intelligence itself. Because
• 1:54
  the Kinshoff dispatch kicks off with this wild
• 1:55
  story about a new kind of IP theft called distillation.
• 2:00
  Yes. Distillation. But before we get into the
• 2:02
  mechanics of this, and I want to make something
• 2:04
  explicitly clear to you listening, we are impartially
• 2:08
  reporting the contents of this source document
• 2:10
  today. Right. That's a very important distinction
• 2:12
  to make. Yeah, because the dispatch details accusations
• 2:16
  made by one company against another. You know,
• 2:20
  we are not endorsing these viewpoints and we
• 2:22
  aren't taking sides here. We are simply conveying
• 2:25
  the claims and the underlying technology that's
• 2:27
  described in the source material. Absolutely.
• 2:29
  And it's an important distinction because, well,
• 2:31
  the claims are explosive. On June 24th, 2026,
• 2:34
  the dispatch notes that Anthropic actually went
• 2:37
  to a U .S. Senate committee. Which is a big deal
• 2:40
  on its own. Right. And they presented a document
• 2:43
  accusing operators tied to Alibaba's Quen Lab
• 2:46
  of executing the largest known distillation attack
• 2:49
  on their clawed models. The numbers they presented
• 2:52
  to lawmakers are just, they're staggering. I
• 2:55
  mean, they alleged roughly 25 ,000 fraudulent
• 2:57
  accounts were created. 25 ,000. Just to bypass
• 3:00
  API rate limits. And through those accounts,
• 3:03
  these operators supposedly ran nearly 29 million
• 3:07
  conversations. Over the course of about six weeks,
• 3:11
  I think. Yeah, six weeks. And the goal here was
• 3:13
  to cheaply copy, or as they call it, distill
• 3:17
  Claude's capabilities, specifically targeting
• 3:20
  its coding, reasoning, and cybersecurity skills
• 3:23
  to train a rival model. Right. Now, Alibaba completely
• 3:27
  denies this, obviously, and as the source notes,
• 3:30
  it hasn't been independently verified. Putting
• 3:33
  the geopolitical drama aside for a second, let's
• 3:36
  talk about the mechanics here. Let's do it. Because
• 3:38
  I understand using bots to scrape data, right?
• 3:40
  That makes sense. But wait, if you're just copying
• 3:43
  the outputs of a model, how does a smaller competing
• 3:47
  model actually learn reasoning? That's a great
• 3:49
  question. Isn't it just like parroting answers
• 3:51
  without understanding the underlying logic? Well,
• 3:54
  what's fascinating here is the underlying mechanism
• 3:56
  of how synthetic data transfer works in modern
• 3:59
  neural networks. You aren't just scraping the
• 4:01
  final answers. When an operator runs a distillation
• 4:03
  attack on this massive scale, they're actually
• 4:06
  prompting the target model to generate what we
• 4:09
  call chain of thought output. Chain of thought,
• 4:11
  right. Right. They are asking the model to show
• 4:13
  its work step by step before arriving at the
• 4:15
  solution. Ah, I see. So instead of just saying,
• 4:18
  you know, write this final line of code. They're
• 4:21
  asking it to explain why it chose a specific
• 4:24
  variable or a specific algorithm in the first
• 4:27
  place. Precisely. To use an analogy, think of
• 4:30
  the frontier model, in this case, Claude, as
• 4:33
  a master teacher. Okay. A master teacher. Yeah.
• 4:35
  A teacher who has already spent years reading
• 4:37
  every single textbook in the world. A distillation
• 4:40
  attack is essentially like forcing that master
• 4:43
  teacher to sit down and write 29 million highly
• 4:47
  detailed step -by -step study guides. Wow. Okay,
• 4:51
  so then the smaller student model. The student
• 4:53
  model doesn't need to read the entire library
• 4:55
  of human knowledge anymore. Right. It just reads
• 4:58
  the study guides. Exactly. It trains directly
• 5:00
  on those pristine, highly structured study guides.
• 5:03
  It learns the logical pathways, the reasoning
• 5:06
  itself, without the massive computational overhead
• 5:08
  of having to figure it all out from scratch.
• 5:10
  That is just wild. I mean, it completely bypasses
• 5:13
  the billions of dollars and, what, months of
• 5:16
  continuous data center compute required for original
• 5:18
  training? Oh, absolutely. You are literally weaponizing
• 5:22
  your competitor's multi -billion dollar machine
• 5:25
  to train your own machine. Which is exactly why
• 5:28
  this is considered asymmetrical warfare. Yeah.
• 5:31
  The cost to defend against this is incredibly
• 5:33
  high. Because think about it, to the frontier
• 5:35
  model servers, this attack just looks like normal,
• 5:39
  albeit heavy, user traffic. Right, because it's
• 5:42
  distributed across thousands of Sybil accounts.
• 5:45
  Just looks like a bunch of users chatting. Exactly.
• 5:48
  And taking this to a U .S. Senate committee signals
• 5:51
  something profound here. It shows that the traditional
• 5:54
  tools we have for protecting intellectual property.
• 5:57
  Like copyrights and patents. Yeah, copyrights,
• 5:59
  patents or even basic terms of service. They
• 6:02
  are completely failing in the AI age. So, OK,
• 6:04
  if software IP is suddenly impossible to defend
• 6:07
  from distillation like your digital moat is basically
• 6:11
  evaporating. Is that why these same companies
• 6:14
  are suddenly pouring billions into proprietary
• 6:16
  silicon? That's exactly it. Are they basically
• 6:18
  trying to build a physical moat because the digital
• 6:21
  one just totally failed? That is the exact strategic
• 6:23
  pivot the entire industry is making right now.
• 6:25
  Wow. If you can't easily lock down the software
• 6:28
  weights of your model, well, the next logical
• 6:30
  step to maintain your edge and your profit margins
• 6:32
  is to aggressively control the physical hardware
• 6:35
  it runs on. And that pivot leads us directly
• 6:37
  to the second major theme in the Kinsoft dispatch,
• 6:41
  which is this. Yes. Two massive stories landed
• 6:44
  on the exact same day and both squarely targeted
• 6:48
  NVIDIA's absolute dominance over the AI silicon
• 6:51
  market. It was a big day. Yeah. First up, OpenAI
• 6:54
  and Broadcom unveiled Jalapeno, which is OpenAI's
• 6:59
  very first custom in -house AI inference processor.
• 7:03
  But here's where it gets really interesting.
• 7:05
  The dispatch reports that this chip went from
• 7:07
  its initial design phase to tape out. In just
• 7:10
  nine months. Right. Nine months. And they're
• 7:12
  planning to deploy it in gigawatt scale data
• 7:14
  centers by the end of the year. I have to push
• 7:17
  back on this timeline a bit. Sure. Nine months
• 7:19
  from design to tape out. That feels like it breaks
• 7:22
  the laws of physics and semiconductor manufacturing.
• 7:24
  Yeah. Even, you know, insanely well -funded teams
• 7:26
  take years to get silicon ready for the foundry.
• 7:29
  Yeah. Usually they do. How is that even mechanically
• 7:31
  possible? Well, it sounds impossible if you assume
• 7:34
  they started from a blank whiteboard. OK. They
• 7:36
  didn't. To pull off a nine month tape out. OpenAI
• 7:40
  heavily leveraged Broadcom's existing pre -verified
• 7:44
  IP blocks. Ah, I see. Think of it like building
• 7:48
  a house with prefabricated walls and plumbing
• 7:50
  rather than mixing the concrete yourself on site.
• 7:54
  Broadcom already has industry -leading SUEs,
• 7:57
  which are these serializer, deserializer blocks
• 7:59
  that handle high -speed data transfer between
• 8:02
  ships. Right, the networking, basically. Exactly,
• 8:05
  and they already have proven memory controllers.
• 8:07
  So OpenAI essentially brought the specific matrix
• 8:11
  math architecture they needed for their AI, and
• 8:14
  Broadcom... just wrapped it in their existing
• 8:16
  foundational tech. So they basically snapped
• 8:18
  these high -end Lego pieces together. Pretty
• 8:21
  much, yeah. But even then, I mean, why go through
• 8:23
  the massive headache of building this when they
• 8:25
  already have access to NVIDIA's top -tier hardware?
• 8:28
  Like, they already have the best stuff. It really
• 8:30
  comes down to the architecture of inference versus
• 8:33
  training. Okay, break that down for me. So...
• 8:36
  NVIDIA's flagship GPUs are incredible general
• 8:40
  -purpose AI accelerators. They are built to handle
• 8:43
  the intense computational precision you need
• 8:46
  for backpropagation during the training phase.
• 8:48
  Right, when the model is actually learning. Yes.
• 8:51
  But Jalapeno is an inference -only chip, meaning
• 8:54
  when a model is just answering user prompts in
• 8:57
  production, you don't actually need that massive
• 9:00
  compute density. What you need at that stage
• 9:03
  is memory bandwidth. You need to move weights
• 9:05
  in and out of SRAM as fast as possible with the
• 9:08
  lowest possible latency. Okay, so it's a completely
• 9:11
  different job. Exactly. By designing a chip strictly
• 9:14
  for inference, OpenAI strips away all that silicon
• 9:17
  real estate dedicated to training and replaces
• 9:20
  it with pure memory throughput. Which means they
• 9:22
  run their models vastly cheaper. faster and cooler
• 9:26
  than using general purpose hardware. Absolutely.
• 9:29
  Okay. It makes total sense for them to carve
• 9:30
  out independence on the physical silicon side.
• 9:32
  But then literally on that exact same day, we
• 9:36
  get the second half of this rebellion. The software
• 9:38
  side. Right. Qualcomm confirmed it's acquiring
• 9:41
  an AI software startup called Modular for over
• 9:44
  $3 .9 billion. A huge acquisition. Massive. And
• 9:49
  the dispatch says Modular created a hardware
• 9:52
  agnostic runtime that attacks NVIDIA's software
• 9:55
  lock -in. Which we know is CDA. Yes, CDA has
• 10:00
  been the gold standard. But again, let's get
• 10:02
  into the mechanics here. People always throw
• 10:04
  around the phrase universal translator for software
• 10:06
  like this. But how does a hardware agnostic runtime
• 10:09
  actually decouple the code from the chip? Well,
• 10:11
  if we connect this to the bigger picture, we
• 10:13
  really have to look at the compiler level. For
• 10:15
  over a decade now, if a developer wrote AI code
• 10:18
  in a framework like, say, PyTorch, the absolute
• 10:21
  easiest way to make it run fast was to use NVIDIA's
• 10:25
  proprietary CUDA libraries because they are highly
• 10:28
  optimized for NVIDIA's specific GPU architecture.
• 10:31
  Right. It created a massive walled garden. You
• 10:34
  basically had to use their hardware if you use
• 10:36
  their software. Exactly. It's a complete ecosystem
• 10:38
  lock -in. Modular tackles this by utilizing an
• 10:41
  intermediate representation layer, specifically
• 10:44
  something called MLIR, or multi -level intermediate
• 10:48
  representation. So it basically sits between
• 10:50
  the high -level Python code and the actual metal
• 10:53
  of the chip. Exactly that. Instead of compiling
• 10:56
  directly to CUDA, Modular's engine takes the
• 10:59
  AI workload, breaks it down into this intermediate
• 11:01
  mathematical representation, and then uses custom
• 11:04
  compilers to translate it perfectly into the
• 11:07
  native instructions. set of an AMD chip, an Intel
• 11:10
  accelerator, or even, you know, a custom chip
• 11:12
  like Jalapeno. Wow. It mathematically decouples
• 11:15
  the software layer entirely from the physical
• 11:17
  hardware layer. So what does this all actually
• 11:19
  mean for the listener? I mean, we have this incredible
• 11:21
  pincer movement happening here. Yeah, a two -pronged
• 11:24
  attack. Jalapeno attacking NVIDIA's hardware
• 11:26
  dominance by proving companies can build custom
• 11:29
  silicon rapidly. And then Modular attacking the
• 11:32
  software dominance by breaking that CDA lock
• 11:35
  -in. Mm -hmm. Why should a business leader or
• 11:38
  just an everyday user care about this highly
• 11:40
  technical silicon war? Because this fierce competition
• 11:43
  is the exact mechanism that will drive down the
• 11:46
  operational costs of the entire digital economy.
• 11:49
  It always comes back to cost. Always. When developers
• 11:52
  aren't forced to buy one specific brand of hardware,
• 11:55
  and when companies like OpenAI aren't paying
• 11:58
  massive premiums for general purpose inference
• 12:00
  compute, the unit economics of AI just... plummet.
• 12:05
  Right. For you, the listener, it means AI integrations
• 12:08
  in your enterprise software, your customer service
• 12:11
  tools, your daily workflows. They will all become
• 12:14
  significantly cheaper and exponentially more
• 12:16
  ubiquitous. You are literally watching the monopolistic
• 12:20
  tax being stripped out of the ecosystem in real
• 12:22
  time. OK, so we are rapidly building out gigawatt
• 12:25
  scale data centers for these custom chips. Yes.
• 12:27
  And we're deploying massive new software compilers
• 12:30
  to route these workloads. Oh. But as we pour
• 12:32
  the concrete for these data centers and write
• 12:34
  millions of lines of new orchestration code,
• 12:36
  the attack surface for bad actors expands massively.
• 12:40
  Oh, without a doubt. And that brings us to the
• 12:41
  final and frankly, the most urgent section of
• 12:44
  the Kinsoft Dispatch today, which is the cybersecurity
• 12:46
  warnings. It is just the unavoidable physical
• 12:49
  reality of infrastructure expansion. You cannot
• 12:52
  scale data centers to the gigawatt level and
• 12:55
  rewrite your entire software stack without introducing
• 12:58
  new vulnerabilities. And the vulnerability highlighted
• 13:01
  in this dispatch is just profoundly ironic. It
• 13:05
  really is. I mean, we just spent 20 minutes talking
• 13:07
  about autonomous reasoning models and hyper advanced
• 13:09
  AI silicon. But on June 25th, the American Cyber
• 13:13
  Agency, CISA, issued an urgent patching directive
• 13:16
  for a phone system. Yep. iPhone system. Specifically,
• 13:21
  Cisco Unified Communications Manager. The bug
• 13:24
  is cataloged as CVE -2026 -2030. It's an unauthenticated
• 13:30
  flaw that let attackers write files directly
• 13:32
  to the underlying system, and it was being actively
• 13:35
  exploited in the wild. Actively exploited, which
• 13:37
  is why the June 28th deadline was so hard. Right.
• 13:39
  But I have to admit some confusion here. Why
• 13:42
  do sophisticated threat actors, who could theoretically
• 13:45
  be targeting these massive AI clusters, care
• 13:47
  about a boring legacy office phone system? Well,
• 13:50
  this raises an important question about network
• 13:52
  topology and something called lateral movement.
• 13:54
  Attackers don't usually try to breach the most
• 13:56
  heavily guarded fortress door, right? They look
• 13:59
  for the forgotten side window. The easy way in.
• 14:01
  Exactly. A unified communications system is essentially
• 14:05
  a server that sits right on the edge of your
• 14:08
  network. It has to connect to the public internet
• 14:10
  to route voice traffic. Sure. But it also has
• 14:13
  deep, trusted connections to your internal enterprise
• 14:16
  network to access things like user directories
• 14:19
  and databases. Ah. So it's basically a bridge
• 14:23
  between... the wild west of the Internet and
• 14:25
  the safe internal corporate network. Yes, exactly.
• 14:28
  And because it's an appliance that just quietly
• 14:30
  works in the background, IT departments often
• 14:33
  treat it as a set it and forget it device. Right.
• 14:35
  It just gets pushed to the bottom of the patching
• 14:37
  queue. Right. And when an attacker exploits an
• 14:40
  unauthenticated file write flaw on that specific
• 14:43
  appliance, they aren't just messing with your
• 14:45
  voicemails. They are dropping a web shell. Oh,
• 14:47
  wow. They gain root access to that server. And
• 14:49
  from there, they pivot. They use that trusted
• 14:52
  appliance to move. laterally inside the network,
• 14:54
  targeting the domain controllers, deploying ransomware
• 14:57
  or, you know, exfiltrating data. Yikes. So the
• 15:01
  grounded practical takeaway for you listening
• 15:03
  is this. You absolutely must treat every single
• 15:06
  piece of internet -facing enterprise software,
• 15:09
  especially the mundane legacy appliances like
• 15:13
  a phone system, as a critical infrastructure
• 15:16
  vulnerability. Which honestly feels a bit overwhelming.
• 15:19
  It is. I mean, the sheer volume of code in modern
• 15:21
  enterprises makes manual patching feel like trying
• 15:24
  to empty the ocean with a teaspoon. Exactly.
• 15:26
  But the dispatch does introduce a massive shift
• 15:30
  in defensive capabilities to counter this, right?
• 15:32
  It does, yeah. Because that same week. OpenAI
• 15:35
  expanded its Daybreak cybersecurity program.
• 15:38
  They released a specialized security -focused
• 15:40
  AI model to vetted defenders. And the pitch here
• 15:44
  isn't just an AI that flags anomalies, right?
• 15:46
  No, it goes much further than that. It includes
• 15:48
  tools for automatically finding and patching
• 15:50
  vulnerabilities in code at scale. So instead
• 15:53
  of a human security analyst endlessly hunting
• 15:55
  for zero days, the AI does static analysis, understands
• 15:59
  the semantic context of the code, and basically
• 16:01
  writes the patch itself. And this brings us to
• 16:04
  the terrifying duality of autonomous agents and
• 16:06
  cybersecurity. Right. Because the concept of
• 16:09
  automated patching, where a system discovers
• 16:11
  a buffer overflow and writes the mitigation code
• 16:14
  in literally milliseconds, it's structurally
• 16:17
  necessary for defenders to keep up at this point.
• 16:20
  Yeah, because humans are just too slow. Right.
• 16:22
  The time to patch window is basically shrinking
• 16:25
  to zero. Yeah. But the dispatch cautions heavily
• 16:28
  about the dual use nature of this technology.
• 16:30
  Right. Because an AI model that is sophisticated
• 16:32
  enough to read millions of lines of code, find
• 16:35
  a deeply hidden logic flaw and understand how
• 16:38
  to patch it. Yes. Is also, by definition, sophisticated
• 16:41
  enough to exploit it. Exactly. We are transitioning
• 16:45
  right now into a machine speed arms race. The
• 16:48
  same static analysis capabilities that empower
• 16:51
  the Daybreak program to secure an enterprise
• 16:53
  can be entirely inverted by threat actors. Wow.
• 16:56
  They can use it to discover zero -day vulnerabilities
• 16:58
  in proprietary software before the vendor even
• 17:01
  knows they exist. That is just wild to think
• 17:04
  about. We are moving away from the era of human
• 17:07
  operators manually probing networks into this
• 17:11
  era where automated AI tools are simultaneously
• 17:13
  defending and attacking infrastructure at speeds
• 17:17
  human analysts simply cannot process. Wow. OK,
• 17:21
  let's take a breath and just bring all of this
• 17:23
  together for you. We started this deep dive looking
• 17:26
  at a single dispatch from June 2026. But it's
• 17:30
  not just a laundry of tech news. It is a high
• 17:33
  definition snapshot. Absolutely. Right. Yep.
• 18:00
  And finally, we looked at the massive security
• 18:01
  fallout of this expansion, balancing the urgent
• 18:04
  reality of lateral network attacks via Cisco
• 18:06
  phone systems against the futuristic machine
• 18:09
  speed promise of OpenAI's Daybreak program. It's
• 18:12
  a lot to take in. And as we consider the architecture
• 18:15
  of this new era, I want to leave you with a final
• 18:18
  thought to mull over, building on that last point
• 18:20
  about security. Okay. We discussed the duality
• 18:23
  of AI and cybersecurity and that shrinking time
• 18:26
  -to -patch window. Well, if AI systems can automatically
• 18:29
  find and patch vulnerabilities at scale, as Daybreak
• 18:32
  promises, what happens when two opposing AIs,
• 18:35
  one designed to relentlessly attack and one designed
• 18:38
  to autonomously defend and engage in an automated
• 18:41
  cyber war at speeds no more than a second, No
• 18:42
  human can monitor. Oh, wow. Will our enterprise
• 18:45
  networks and the basic Internet infrastructure
• 18:47
  we rely on every day simply become passive battlefields
• 18:51
  for autonomous agents fighting invisible millisecond
• 18:53
  wars? That is a wild and honestly somewhat chilling
• 18:56
  thought to leave on. The landscape is entirely
• 18:59
  new and the fundamental infrastructure of the
• 19:02
  digital world is being rewritten as we speak.
• 19:04
  We hope this deep dive gave you the mechanistic
• 19:07
  context you need to truly understand these shifts,
• 19:09
  because taking the time to explore the how and
• 19:11
  why is exactly how you stay ahead of the curve.
• 19:14
  Thanks for joining us today. Stay curious and
• 19:16
  stay patched.