# GPT-6 Astra Patched Qwen’s Math—But the Same Fix Failed on Gemma

Canonical URL: https://superpowerdaily.com/videos/fd963481-8df7-4ec2-ba4f-039bfcb9e157
Published: 2026-11-03T09:00:16.000Z
Episode date: 2026-11-03
Language: English
Duration: 9:49
Publisher: Superpower Daily
YouTube: https://www.youtube.com/watch?v=4Lx1Aj_BsXc

MIT researchers used GPT-6 Astra to surgically edit Qwen’s weights, improving its math performance, though the approach failed on Gemma. OpenAI rebuilt Git for Codex-driven parallel work, experts infer North Korean hackers are using AI to target crypto firms, and Anthropic’s temporary message threads speed Claude agent requests without reducing token charges.

## Key takeaways

- A targeted weight edit reduced Qwen's repetition and boosted
- Pando lets unrelated branches advance independently, but
- Investigators infer AI use in crypto targeting from attack
- Message threads reduce upload overhead, not token charges;

## Chapters

- 0:31 Targeted model patches: https://superpowerdaily.com/videos/fd963481-8df7-4ec2-ba4f-039bfcb9e157?t=31
- 3:00 Parallel code infrastructure: https://superpowerdaily.com/videos/fd963481-8df7-4ec2-ba4f-039bfcb9e157?t=180
- 4:57 AI and crypto threats: https://superpowerdaily.com/videos/fd963481-8df7-4ec2-ba4f-039bfcb9e157?t=297
- 6:48 Lighter agent requests: https://superpowerdaily.com/videos/fd963481-8df7-4ec2-ba4f-039bfcb9e157?t=408

## Stories covered

- MIT researchers use AI to patch a model, raising its math score from 7.9% to 48.1%: https://superpowerdaily.com/posts/mit-researchers-use-ai-to-patch-a-model-raising-its-math-score-from-7-9-to-48-1
  GPT-6 Astra’s direct edits mainly stopped Qwen from repeating itself. Finding the tiny patch took substantial investigation, and the researchers did not demonstrate self-improving AI.
- OpenAI rebuilds its Git backend to handle more parallel work from Codex: https://superpowerdaily.com/posts/openai-rebuilds-its-git-backend-to-handle-more-parallel-work-from-codex
  Pando now powers development in OpenAI’s shared code repository. Its design lets unrelated branches advance independently, but gives up a single synchronized view across them.
- North Korean hackers use AI to hunt high-value crypto targets, experts say: https://superpowerdaily.com/posts/north-korean-hackers-use-ai-to-hunt-high-value-crypto-targets-experts-say
  The assessment centers on TraderTraitor’s targeting methods and recent attack techniques—not a demonstrated account of AI carrying out an entire theft.
- Anthropic adds message threads so Claude agents stop resending full conversations: https://superpowerdaily.com/posts/anthropic-adds-message-threads-so-claude-agents-stop-resending-full-conversations
  The public beta reduces repeated uploads rather than the conversation Claude reads. Anthropic reports faster starts for long requests, but developers must still keep their own history.

## Transcript

Automatically transcribed from the narration.

[0:00] Welcome to today's superpower daily AI briefing explainer We've got a really exciting one for you today Translating four critical AI developments from November 3rd 2026 into clear actionable insights We have a packed briefing so let's just jump right in Over the next few minutes we're gonna be covering MIT's wild breakthrough in literally patching an AI model OpenAI's massive Git backend rebuild How North Korean hackers are weaponizing AI to go after crypto And finally Anthropic's brand new

[0:29] message threads for their cloud agents All right section one MIT AI patches the Quen model directly editing open weights to boost math performance So picture this The MIT team takes the Quen 3.5 0.8 B model right And it is just bombing the math 500 test scoring a dismal 7.9% But then they used OpenAI's GPT 6 Astra to effectively patch it The result That score absolutely skyrocketed to an impressive 48.1% Now what's mind blowing here is that

[1:03] Astra didn't teach Quen new mathematical concepts It didn't run it through more traditional training data Instead it acted like a highly specialized neurosurgeon It diagnosed an internal obstruction basically a bottleneck that was causing the model to just endlessly repeat itself until its context window was totally full leaving it no room to actually generate the right answer And Astra just removed that obstruction To really grasp how crazy this is we got to look at the scale Astra

[1:30] pulled off this massive performance jump by changing a microscopic 0.016% of the model To put that into perspective we're talking about just 131,072 weights the internal numerical parameters located specifically in layers 15 and 19 of the neural network By targeting those exact spots Astra surgically suppressed a specific numerical direction that was associated with all that repetition You know when a massive AI model gets stuck in a loop it's just spinning its wheels wasting all its computational

[2:01] power By simply turning off that repetitive tendency Astra freed the model up to use the math skills it actually already possessed right there under the hood But and this is a big but you have to know this wasn't some instant magical fix It actually took 20.3 hours 158 conversational turns and 80 Python runs to nail down this exact precise edit Plus when the researchers tried to apply this exact same methodology to a totally different model the

[2:30] GEMMA 3270M it completely failed to improve its math scores Why Because GEMMA just didn't suffer from that specific repetition bottleneck in the first place Its architecture had totally different weaknesses So the big takeaway here while this is absolutely incredible detective work it doesn't prove that AI can recursively self improve in a reliable continuous cycle across different architectures just yet Right now it remains a highly isolated manual surgical fix rather than true autonomous evolution OpenAI rebuilds their

[3:02] Git backend Pando handles codex driven parallel activity Okay let's meet Pando This is a custom built Git backend engineered by OpenAI to serve their massive shared codebase which in the developer world is called a monorepo The core issue driving this was that their standard Git infrastructure was literally buckling under the unprecedented pressure of parallel tasks generated by codex which is OpenAI's coding model Think about it You've got human engineers and thousands of AI agents pushing pulling

[3:32] and merging code simultaneously 24 hours a day Every single action triggers this avalanche of automated checks and it completely overwhelmed their traditional servers They desperately needed a fundamental rebuild from the ground up but crucially they needed one that wouldn't force their human engineers to abandon the standard Git workflows and commands they already knew and used every day By late last year that old system was visibly struggling to keep up with this hybrid human AI workforce Code

[3:59] fetches slowed to a total crawl merges were routinely failing and overall system reliability tanked below 90% Pando fixes this severe bottleneck by doing something really clever It intelligently isolates branch updates Instead of making every single code change wait in one giant shared queue to be processed Pando lets unrelated branches advance completely independently It also highly optimizes data transfer by reusing compressed bundles of code called pack files sending identical bytes to a bunch of different readers all

[4:29] at once This elegant solution actually cuts push latency by nearly five times letting the system serve almost a petabyte of Git data every single day The sacrifice they had to make for all that speed Pando relies on a concept called causal consistency For an engineer this practically means you get the accurate history you need for your specific branch but the backend system no longer guarantees one perfectly synchronized global snapshot of every single branch across the entire

[4:54] repository at the exact same millisecond Let's shift gears to section three hackers target crypto with AI North Korea's Trader Trader group evolves big game hunting So cybersecurity experts like Park Seong su at Scalar are sounding the alarm that cyber operations are accelerating dramatically both in speed and complexity In fact the FBI attributes a staggering $6.3 billion Binance theft directly to this specific North Korean group called Trader Trader Researchers are calling their evolving strategy big game hunting

[5:28] Instead of wasting time targeting individual retail investors with standard phishing scams they are highly suspected of using AI to rapidly study a target company's entire infrastructure and internal fund transfer processes The AI basically acts as an ultra fast scout mapping out vulnerabilities so they can execute massive single blow thefts against high value targets Now to be totally clear security experts haven't necessarily seen an AI push the final button on the theft itself but they are strongly

[5:56] inferring deep AI adoption based purely on how highly accelerated and sophisticated these new reconnaissance and attack techniques have become In the sheer scale of the automation they use to hide the stolen money after the breach it is absolutely mind boggling Investigators at Elliptic noted that the laundering activity for these specific attacks moved through about 12,000 distinct wallets across 11 different blockchains And they did all of this in under four days If you break that math down

[6:24] that is roughly one new wallet generated every 25 seconds consistently for days on end This is a level of speed coordination and complexity that practically guarantees the use of advanced AI driven automated systems There is simply no way human operators could create wallets breach assets and execute complex cross chain swaps at a sustained rate of one every 25 seconds without making a single mistake Finally section four Anthropic launches message threats Clawed agents stop resending full conversations

[6:55] If you've ever built an AI coding agent you know the friction of how it currently works in practice Without message threats every single time an AI agent makes a request to the API it has to upload the entire history of the conversation all the relevant code based context all previous tool results the whole shebang over and over again For developers building these agents that causes massive delays huge bandwidth bottlenecks and incredible frustration during long coding sessions

[7:22] Well now clawed agents can execute a much simpler workflow You just send that massive initial payload of conversation system instructions and available tools exactly once Then on later turns Anthropic temporarily stores that context so the agent only has to append the brand new messages along with the previous response ID It's really crucial to understand what this update actually fixes for developers and what it definitely doesn't Yes it massively improves what we call time to first token

[7:50] meaning Clawed starts typing its answer way faster because the payload traveling over the network is significantly smaller However Clawed is still reading and processing that full conversation on its end every single time it generates a response So your token charges absolutely do not decrease Plus these threads are highly temporary They expire just five minutes after the last request or up to an hour at the absolute maximum And because the backend storage is considered best effort if

[8:18] a continuation fails for any reason you've got to be prepared to replay that entire history yourself to recover the threat Ultimately this is an infrastructure update built entirely to solve for speed and lower latency in agent workflows not to reduce your monthly token bill All right let's pull all this together into section five four takeaways and closing summarizing today's AI briefing To do a quick recap of today's explainer number one MIT researchers proved you can surgically

[8:45] patch an AI model's behavior directly in its weights to clear bottlenecks though it does require hours of testing and remains a targeted fix not a magic bullet for broad self improvement Number two OpenAI's Pando project highlights that the massive infrastructural demands of AI coding agents require entirely new parallel backend paradigms that leave traditional Git limitations in the dust Number three state sponsored hackers are now leveraging AI to automate their reconnaissance and execute crypto laundering at unprecedented

[9:15] machine speed scales And number four Anthropic's new message threads offer a much needed temporary fix for heavy API payloads speeding up response times even if they don't lower your token costs With all of these rapid advancements happening across the board we have to ask a broader question As AI aggressively pushes the boundaries of our internal infrastructure our model architectures and our security defenses which of these critical bottlenecks is going to break first Read more at superpowerdaily

[9:44] com Stay curious

## Editorial standards

https://superpowerdaily.com/editorial-standards

This episode reports information available at publication time. Consult the linked stories and their reporting sources for context and subsequent updates.