Latest / AI Fire Daily / 🎙️ EP 390: MIT's Weight-Space Programming Breakthrough & DeepSeek’s Super-Agent!

Transcript

🎙️ EP 390: MIT's Weight-Space Programming Breakthrough & DeepSeek’s Super-Agent! · AI Fire Daily

• 0:00
  Imagine waking up tomorrow. You just perform
• 0:03
  laser eye surgery on yourself. Yeah, that sounds
• 0:06
  crazy. Right. You do this without any medical
• 0:08
  training whatsoever. It sounds completely impossible,
• 0:11
  but overnight, an AI basically did exactly that.
• 0:15
  It performed surgery on another AI's brain. It
• 0:17
  is a genuinely terrifying and, you know, incredible
• 0:20
  concept. We are watching these systems manipulate
• 0:23
  their own core logic. Welcome to the Deep Dive.
• 0:27
  Today, we are exploring a massive leap in AI
• 0:29
  autonomy. We are tracking a really fascinating
• 0:32
  journey through this new landscape. First, we
• 0:34
  will look at AI directly editing its own underlying
• 0:37
  code. Then we will look at super agents coordinating
• 0:41
  entire parallel workflows. And finally, we will
• 0:44
  examine the massive costs of this new independence.
• 0:47
  The ground is shifting incredibly fast right
• 0:50
  now. I mean, this is no longer just about generating
• 0:52
  a nice email. It is about autonomous action at
• 0:55
  a fundamental... structural level. So we begin
• 0:58
  at the microscopic level today. We need to look
• 1:01
  at the very architecture of an AI's brain. MIT's
• 1:05
  CSAIR researchers just published some incredible
• 1:08
  new findings. They introduced a technique called
• 1:10
  weight space programming. This is a massive departure
• 1:14
  from how things usually work. Traditionally,
• 1:17
  improving a model requires massive amounts of
• 1:19
  new data. You basically have to retrain the whole
• 1:22
  thing from scratch. Right. It is incredibly expensive.
• 1:25
  and extremely time -consuming. But this new technique
• 1:28
  changes that dynamic entirely. They used a model
• 1:31
  called GPT -6 Astra. Astra directly read and
• 1:35
  edited another neural network. This was QN3 .50
• 1:38
  .8b. Yeah, and it did this without any traditional
• 1:41
  training. It just reached right in and adjusted
• 1:45
  the weights directly. For anyone unfamiliar,
• 1:48
  we should quickly define what weights actually
• 1:50
  are. Right. Think of weights as the digital brain
• 1:53
  connections determining how an AI thinks. Exactly.
• 1:56
  They are the structural pathways of the model.
• 1:59
  And Astra identified a highly specific problem
• 2:03
  inside Quinn's architecture. The smaller model
• 2:06
  was getting stuck in a bizarre reasoning loop.
• 2:08
  Uh -huh. So Astra located the flaw, then it actually
• 2:12
  fixed it using Python. Instead of retraining
• 2:15
  the whole model with new data. Right, which is
• 2:17
  basically like teaching a student math completely
• 2:19
  from scratch. Astra mapped the weight space to
• 2:22
  locate the exact coordinates. They found the
• 2:25
  bad logic circuit and simply rewired it. The
• 2:27
  statistical improvements from this surgery are
• 2:29
  just staggering. Let's look at the Math 500 accuracy
• 2:33
  scores. Quinn was originally scoring a pretty
• 2:36
  dismal 7 .9%. Right. After Astra's direct edit,
• 2:40
  it jumped to 48 .1%. That is a six -fold improvement.
• 2:44
  In an instant. The GSM 8K math scores are equally
• 2:47
  wild to look at. Quen went from 13 .8 % up to
• 2:51
  67 .1%. Its ability to follow instructions also
• 2:56
  tripled to 62 .3%. Right. And here is the most
• 3:00
  mind -blowing detail. Astra only modified a microscopic
• 3:03
  fraction of the actual network. Oh, yeah. It
• 3:06
  changed just 0 .016 % of Quen's total weights.
• 3:15
  That is like finding a needle in a haystack of
• 3:18
  needles. It knew exactly which microscopic connections
• 3:20
  were causing the failure. It is like stacking
• 3:24
  Lego blocks of data, but the AI is replacing
• 3:27
  the foundational blocks. And it does this without
• 3:30
  knocking over the tower. That is a perfect analogy
• 3:33
  for what is happening here. It is highly targeted,
• 3:35
  microscopic digital surgery on the fly. They
• 3:38
  also tested this on gaming models with wild results.
• 3:42
  Astra could identify four different video games
• 3:45
  just by looking. It did not watch the games being
• 3:48
  played on a screen. Wow. It looked directly at
• 3:50
  the raw neural network weights. It identified
• 3:53
  games like Pong and Breakout instantly. It just
• 3:56
  read the matrix of raw numbers. Then it fixed
• 3:58
  a game -playing AI's completely broken behavior.
• 4:01
  It took its success rate from 0 to 100%. But
• 4:08
  we have to acknowledge the current limitations
• 4:09
  here. This is not a magic wand for every single
• 4:12
  model yet. The researchers tried this exact technique
• 4:15
  on a smaller GEMMA model. Yeah, and it could
• 4:17
  not significantly improve that GEMMA model. The
• 4:20
  GEMMA model actually lacked that specific reasoning
• 4:22
  loop problem. Because there was no localized
• 4:25
  flaw, Astra had nothing to patch. The research
• 4:29
  is still very preliminary, though logs are fully
• 4:32
  open. But I really have to push back on this
• 4:35
  a bit. If this only works for highly specific
• 4:38
  localized reasoning loop problems, if it needs
• 4:40
  a clear target to actually go in and fix, isn't
• 4:44
  this just a targeted bug patch rather than true
• 4:47
  learning? Well, it is a patch job right now,
• 4:49
  yes. But think about the broader implication
• 4:52
  of what this means. I mean, the hardest part
• 4:54
  of software engineering isn't writing the code.
• 4:57
  Right, it's debugging. Exactly. The hardest part
• 5:00
  is actually debugging the code. If an AI can
• 5:03
  autonomously hunt down a microscopic logic flaw,
• 5:06
  if it can find it in the billion parameter haystack
• 5:08
  and patch it, we are inches away from true self
• 5:13
  -compounding intelligence. Hmm. What happens
• 5:17
  when AI starts doing this recursively? What happens
• 5:21
  when it fixes its own bugs in real time? That
• 5:24
  is the million dollar question for the entire
• 5:26
  industry. On one hand, you get an incredible
• 5:29
  speed of self -improvement. These models could
• 5:32
  evolve faster than human engineers ever could.
• 5:35
  But on the other hand, you face the danger of
• 5:37
  runaway errors. If it makes a bad edit autonomously,
• 5:41
  that error cascades rapidly. It could corrupt
• 5:44
  its own underlying logic entirely without human
• 5:47
  oversight. So self -correcting AI speeds up evolution,
• 5:51
  but desperately needs strict oversight. Exactly.
• 5:55
  It is a massive structural leap, but a precarious
• 5:58
  one. If a single AI can autonomously edit its
• 6:02
  own brain, the natural next step is for these
• 6:04
  models to team up. What happens when you connect
• 6:07
  multiple autonomous brains together? Yeah. What
• 6:10
  happens when they orchestrate massive real -world
• 6:13
  projects in tandem? You step into a parallel
• 6:15
  universe of autonomous coordination. And DeepSeek
• 6:19
  just gave us a very clear window into this. They
• 6:21
  recently open -sourced something called the DSH
• 6:23
  superagent prototype. Right. This is a system
• 6:27
  that coordinates multiple coding agents simultaneously.
• 6:30
  They do not work sequentially waiting for one
• 6:33
  another anymore. They handle highly complex tasks
• 6:36
  entirely in parallel. They use a central router
• 6:40
  to assign all the work. This router essentially
• 6:42
  acts like a digital project manager. It lets
• 6:45
  multiple subagents operate completely independently
• 6:48
  on different tasks. These agents monitor live
• 6:52
  digital environments for new assignments. They
• 6:54
  watch GitHub issues and pull requests constantly
• 6:57
  without sleeping. They even monitor FaceYou corporate
• 7:00
  chats for incoming instructions. They do the
• 7:02
  requested work and return the results autonomously.
• 7:04
  They can even submit their own completed pull
• 7:07
  requests. Wow. The whole ecosystem is tracked
• 7:10
  via a simple web dashboard. The system only asks
• 7:13
  for human input when absolutely needed. The early
• 7:16
  test results on this are pretty staggering. They
• 7:19
  completed complex software tasks in just 12 .4
• 7:22
  minutes. A standard single session baseline took
• 7:25
  20 .4 minutes for that same work. Right. And
• 7:28
  it drastically reduced the worst case scenario
• 7:30
  times. The 95th percentile completion time dropped
• 7:33
  massively during these tests. It went from 29
• 7:36
  .5 minutes down to just 15 .7 minutes. But giving
• 7:42
  an AI this much control is inherently risky.
• 7:45
  I still wrestle with prompt drift myself. Oh,
• 7:48
  that happens to absolutely everyone testing these
• 7:50
  systems. Prompt drift is just a frustrating reality
• 7:53
  right now. For those wondering, prompt drift
• 7:56
  is exactly what it sounds like. It is when an
• 7:58
  AI hallucinates entirely new instructions mid
• 8:00
  -task. I will give an AI a simple set of coding
• 8:04
  instructions. Ten minutes later, it is doing
• 8:06
  something completely different. If humans struggle
• 8:10
  to keep a single session on track, how do we
• 8:13
  blindly... trust a central AI router? How do
• 8:16
  we trust it managing an entire swarm of agents?
• 8:19
  Well, it requires an entirely new framework for
• 8:22
  systemic control. You are not managing a single
• 8:25
  digital tool anymore. You're managing an entire
• 8:27
  digital workforce acting autonomously. Imagine
• 8:31
  scaling to a billion queries. Right. The level
• 8:35
  of orchestration required behind the scenes is
• 8:37
  just hard to fathom. How do we prevent these
• 8:40
  swarms from running in circles? How do we stop
• 8:42
  them from burning through computing power? You
• 8:45
  have to establish incredibly tight parameter
• 8:46
  controls up front. You need hard, unbreakable
• 8:50
  limits on how many loops they run. But fundamentally,
• 8:53
  you are accepting a major operational tradeoff
• 8:55
  here. You accept higher token usage in exchange
• 8:58
  for pure parallel speed. Basically, we are trading
• 9:02
  a massive spike in token costs for faster execution.
• 9:05
  Yeah, that is the reality of the current autonomous
• 9:07
  architecture. Speed and autonomy cost serious
• 9:10
  money in the form of tokens. DeepSeek actually
• 9:13
  warns that this is an experimental research prototype.
• 9:16
  It comes with potential security risks and substantial
• 9:19
  bugs right now. That massive spike in token costs
• 9:23
  is not just theoretical. It is happening right
• 9:26
  now to real people and major companies. They
• 9:29
  are building out this autonomous infrastructure
• 9:31
  as we speak. The financial realities of autonomous
• 9:34
  AI are finally hitting home. We are witnessing
• 9:37
  the death of AI as traditional software. We don't
• 9:40
  just buy it once and own it anymore. Look at
• 9:43
  Gary Tan, the CEO of Y Combinator. He reportedly
• 9:45
  spends up to $200 ,000 a year on AI tokens. Which
• 9:49
  is just wild. He uses these autonomous agents
• 9:52
  for coding and daily life tasks. That is an unbelievable
• 9:55
  amount of money for individual usage. But he
• 9:59
  is deeply integrated into this bleeding edge
• 10:01
  ecosystem. He even recently open sourced 23 cloud
• 10:05
  code skills in GetStack. He's freely sharing
• 10:08
  the tools that burn through those tokens. OpenAI
• 10:11
  is reacting to this massive surge in heavy usage.
• 10:14
  They are making a huge structural change this
• 10:17
  coming December. They are merging chat GPT and
• 10:20
  codecs into one shared usage pool. This applies
• 10:22
  to all paid plans moving forward. Thankfully,
• 10:25
  they stated instant chats will not count towards
• 10:28
  limit, but developers are going to have to spend
• 10:30
  tokens very carefully. The era of unlimited autonomous
• 10:33
  looping is getting extremely expensive. We are
• 10:36
  also seeing new tools emerge to streamline this
• 10:39
  chaos. A SpaceX AI engineer recently created
• 10:42
  something called Dr. Eggbot. It is a plug and
• 10:44
  play bot that literally creates other bots. It
• 10:47
  builds better grok bots automatically based on
• 10:49
  specific needs. It is autonomous infrastructure
• 10:52
  building its own optimized infrastructure. We
• 10:55
  are also seeing platforms like beacon .sh stepping
• 10:58
  up significantly. They provide shared, self -improving
• 11:01
  memory for these autonomous tools. Beacon .sh
• 11:04
  works with Cloud Code, Cursor, and Codex directly.
• 11:07
  It helps these agents get smarter with every
• 11:10
  single session. Which means they use fewer cached
• 11:12
  tokens over time. Cached tokens are pre -processed
• 11:16
  text chunks that save time and computing power.
• 11:18
  By remembering context, the AI doesn't have to
• 11:21
  reprocess everything. Spending $200 ,000 on tokens
• 11:26
  reveals a massive paradigm shift. We are moving
• 11:29
  away from treating AI like software we buy. We
• 11:32
  are treating it like a utility bill we keep running.
• 11:34
  It is exactly like paying for electricity or
• 11:37
  running water. The more agents you run, the higher
• 11:40
  the monthly utility bill. Will everyday users
• 11:43
  ever benefit from this level of automation? Or
• 11:46
  is it priced strictly for wealthy tech CEOs?
• 11:49
  right now it is definitely skewed heavily toward
• 11:51
  the high end but tools like beacon .sh will eventually
• 11:55
  drive efficiency way up that optimization will
• 11:58
  inevitably drive the costs down for regular users
• 12:01
  yes once shared memory tools make these autonomous
• 12:04
  agents much more efficient exactly it'll just
• 12:07
  take time for the architectural optimization
• 12:09
  to trickle down let's take a quick moment to
• 12:12
  thank our sponsors welcome back to the deep dive
• 12:19
  With AI acting autonomously and chewing through
• 12:21
  massive utility budgets, governments and developers
• 12:24
  are rushing to establish strong enforceable guardrails.
• 12:27
  They desperately need to track what is real and
• 12:30
  what isn't. They also need to see if this autonomy
• 12:32
  actually works for humans. We are seeing a massive
• 12:36
  wave of new benchmarks and regulations. OpenAI
• 12:38
  just launched a brand new benchmark called LifeVal.
• 12:41
  It is built specifically to test actual everyday
• 12:44
  human tasks. We are finally moving past abstract.
• 12:48
  highly theoretical math problems. LifeVal tests
• 12:51
  things people actually care about in their daily
• 12:53
  reality. It evaluates how agents handle childcare
• 12:57
  scenarios and personal finance. It even tests
• 13:00
  how they navigate complex government service
• 13:02
  portals. We are also seeing new models built
• 13:05
  to predict human behavior. A company called 42AI
• 13:08
  just raised $62 .5 million. They are building
• 13:12
  what they call a behavioral world model. It is
• 13:15
  designed to estimate how people react in new
• 13:17
  situations. They are already testing this model
• 13:20
  with major financial brokerages. It tries to
• 13:23
  simulate market reactions and human panic before
• 13:26
  they happen. We are seeing massive scale in creative
• 13:29
  and business workflows too. Longcat 2 .5 just
• 13:32
  launched with 1 .6 trillion total parameters.
• 13:36
  It is built specifically for complex coding and
• 13:38
  advanced 3D projects. And businesses are starting
• 13:41
  to deploy these autonomous systems directly.
• 13:43
  Unleash X is a platform letting businesses literally
• 13:47
  hire AI employees. These agents handle real workflows
• 13:50
  with humans involved only when needed. But I
• 13:54
  want to circle back to 42AI for a second. It
• 13:57
  feels mildly dystopian when you really think
• 14:00
  about the mechanics. Oh, for sure. What happens
• 14:02
  when a financial brokerage uses a behavioral
• 14:04
  world model? What happens when they predict exactly
• 14:07
  what makes you panic sell? That is the incredibly
• 14:11
  dark side of predictive behavioral modeling.
• 14:14
  If they can accurately model your fear, they
• 14:17
  can monetize your fear. Wow. It is exactly why
• 14:20
  regulation is becoming the dominant conversation
• 14:22
  right now. The EU AI Act is already forcing major
• 14:26
  changes across the industry. Invisible watermarking
• 14:29
  is rapidly becoming an enforced industry standard.
• 14:32
  Starting November 7th, SpaceX is officially adding
• 14:35
  it to Grok. All Grok -generated API techs will
• 14:38
  carry this invisible trackable watermark. Anthropic
• 14:41
  is doing the exact same thing to meet EU rules.
• 14:44
  They are embedding invisible text across their
• 14:46
  major frontier models. That includes Claude Mythos
• 14:49
  5, Opus 4 .7, and Opus 4 .6. It also includes
• 14:54
  Sonnet 4 .6 and all of their newest iterations.
• 14:57
  But do these invisible watermarks actually stop
• 15:01
  the misuse of AI? Or are they just a superficial
• 15:04
  feature to keep regulators happy? Honestly, they
• 15:07
  are mostly just a basic compliance tool today.
• 15:10
  Sophisticated bad actors will inevitably find
• 15:12
  ways to strip them out completely. But it is
• 15:15
  a necessary first step for establishing legal
• 15:17
  accountability. It is mostly a compliance checkbox
• 15:20
  that will not stop sophisticated bad actors.
• 15:23
  Yeah, but regulators have to start somewhere
• 15:26
  to build the framework. We have covered an immense
• 15:30
  amount of ground today. We really have. If you
• 15:33
  look at the logical through line of everything
• 15:34
  we discussed. We have watched AI graduate from
• 15:37
  a simple, prompted response tool. It is now something
• 15:41
  that can microscopically fix its own internal
• 15:43
  logic. It can manage entire teams of subordinate
• 15:46
  bots in parallel workflows. And it can cost hundreds
• 15:50
  of thousands in monthly utility fees. The sheer
• 15:52
  scale and speed of this autonomy is truly breathtaking.
• 15:57
  It really forces you to reevaluate where this
• 15:59
  is all heading. I want to leave you with one
• 16:02
  final thought today. Think about that laser eye
• 16:05
  surgery metaphor from the very beginning. Right.
• 16:08
  The AI reached inside itself and rewrote its
• 16:12
  own foundational code. It bypassed human engineers
• 16:15
  entirely to become significantly smarter. Now
• 16:18
  combine that capability with tools like 42AI's
• 16:22
  behavioral model. If AI can perfectly model your
• 16:25
  daily choices, habits, and fears, and it can
• 16:30
  seamlessly rewrite its own code to predict you
• 16:32
  better, At what point are we the ones actually
• 16:35
  being programmed? That is a deeply unsettling
• 16:38
  yet entirely possible reality to consider. Thank
• 16:41
  you for joining us on this deep dive. Take a
• 16:43
  close look at your own daily workflows this week.
• 16:46
  Examine how many everyday repetitive tasks you
• 16:49
  could autonomously automate today. You might
• 16:53
  be completely surprised by what is already technically
• 16:55
  possible. Stay curious and we will see you next
• 16:58
  time.