5 insights, human curated, human summarized and optimized for busy AI engineers. Here is the high signal/noise ratio news since the last issue.
Alex: this issue took a bit longer than usual because per my commitment to you, I was waiting for 5 genuinely top quality insights instead of wasting your attention.
A lot has happened this month: from the Nvidia CEO scandal to Anthropic IPO but to me, the most important news are direct weight editing techniques. These are solutions to a key problem I called out in Coding Is Not Solved: the ability to learn after the training cut-off.
1. LLMs can natively read & write neural network weights
Rikiya Takehi, et al., 2026-11-02
Problem: Humans don’t understand model weights and training them consumes a lot of compute and requires extensive evals.
Solution: According to this MIT research models like Astra can directly tweak them without a runtime! It’s a well written article with lots of visuals. You should check it out.
Key takeaways:
LLM could make an SLM better at math
LLM could beat a model that was trained in 100M steps by simply editing the weights
There are some caveats and counter-arguments that came out one day later.
Alex: the model weights are akin to organic DNA. This approach is the CRISPR moment of AI because for the first time (to my knowledge) we can create or edit a feature instead of “growing it”. They call it “Weight-Space Programming” which treats the model’s weights as a programming language.
This is a key idea for RSI (recursive self improvement) and despite having huge potentials (e.g. making a Small Language Model more capable) it has huge risks too because our ability to understand and control AI rapidly dip below the level of their improvement.
A few related links:
Hope AI by Google
Language Models Can Rewrite Their Own Weights (came out 2026-11-16)
2. Amazon saves millions by replacing Sonnet 4.6 with Strands Decider
Strands, 2026-11-09
Problem: Amazon was using Claude Sonnet 4.6 for labeling internal tickets.
Solution: Decision models could do the job 78% faster and 98% cheaper
Key takeaways:
This was a success because the catalog doesn’t change that often, the volume is high and the latency matters.
Use shadow loading (launching both systems to prod and observe their diff) to gather data on accuracy, latency, and cost. The next best thing is feature-flag. Regardless, you may want to keep both systems (one as a fallback) if you can prevent feature drift.
Alex: back in October I wrote an article about AI hype and I used Jev as a masterfully executed example. Many critics (who didn’t read the actual article) thought I’m bashing Jev. It couldn’t be further from the truth. Decision model is a new architectural construct and this is a good example of where they’re the better solution.
3. What comes after AGI?
Google DeepMind, 2026-11-19
Idea: AGI is just one point in a spectrum that contains ASI (artificial superintelligence) and “Universal AI”.
Key takeaways:
Thanks to improved algorithms, it is 3x-6x cheaper to train a new model. In other words, we have 3-6 times more compute at our disposal without building or buying new hardware.
The article then extrapolates other improvements to conclude: “Multiplying all three factors gives an overall growth rate of global effective compute stock of at least 10x per year - a factor of 10 billion over ten years”
Alex: it’s an interesting read and shows how an AI lab thinks about compute, the future of humanity, data exhaustion, diminishing returns and growth gaps.
Many ideas in this article sound like they concern a distant future but a key point in the article makes them closer than they are: multiple paths to RSI (recursive self improvement) which Robert Miles, the AI safety YouTuber, calls to ban (YouTube).
4. Lovable Apps Expose Sensitive User Data
crissyfield, 2026-11-02
What happened?
Lovable chose Supabase but implemented it incorrectly.
Supabase has a relatively lean Row Level Security policy by design (here’s how to secure it)
Sensitive data were accessible without logging in:
A German lifestyle forum exposed +80k profiles with PII and +200k chat messages
Tens of thousands of event organizing records
+1000 records containing identity records
The accountability gap:
Lovable marketing repeatedly claims that “No technical knowledge is needed to build projects” and employees proudly present themselves as “building the last piece of software” on LinkedIn while their CEO explicitly states: “We’re building for the nontechnical and the 99%”
But their terms & conditions explicitly holds the [same non-technical target] customers accountable for “reviewing, validating, and using any AI Output” and “your Customer Data”. Lovable explicitly states: “you agree not to upload, input, or otherwise provide […] special or sensitive categories of data (including financial account numbers, payment card data, government identifiers, or biometric data)” which is legally sound but brings the utility of their service to toy level POC.
Lovable has an automated scanner that requires user action. There’s also an auto-fix feature according to Igor, head of Security at Lovable (12 days after the report above).
Alex: This is a good example of how the platform acts as a multiplier and how AI lowers the entry bar while simultaneously raises the bar for doing it right.
I live in Stockholm, the birthplace of Lovable, and had my share of interactions with the crew behind the service. What I observed was a lot of young energy and enthusiasm in a company that’s growing rapidly thanks to Nvidia’s massive investments in a bunch of startups to [artificially] inflate GPU demand.
Lovable has been struggling with security and reliability as long as I can remember (another example) and unfortunately I believe it starts from their talent intake process explicitly filtering for velocity over quality or depth of experience. So when I saw the news I wasn’t surprised a bit. Good luck to Igor and colleagues securing Lovable because as I have elaborated in my book, these are socio-technical concerns. You cannot separate the reliability and security of a system from the people and culture behind it.
5. AI Productivity Fallacies
Alex Ewerlöf, 2026-11-09
What: This is my own post debunking a few productivity fallacies around AI
Key takeaways:
To Managers: AI has a low entry bar but a high bar for productivity. Instead of stressing and threatening your engineers, support them and trust their judgment.
To Engineers: Stand your ground on the quality vs velocity trade-off because the AI cannot be held accountable.
Alex: this came out of my viral post on social media so I decided to add some more meat to the discussion and publish it as a post that digs deeper with pragmatic advice for both managers and engineers. If you prefer the video format, here’s one about the Harvard study showing that while the output increased we don’t necessarily have better outcomes (YouTube)
Out of scope, but interesting
These are some links that didn’t make it to the top-5 due to my strict criteria and promise to you. If you have time, have a look:
2026-11-18: Someone fooled AI by going viral based on misinformation (YouTube short)
2026-11-14: Weight Space Misfolding, a technique for modifying a neural network’s knowledge, behavior and alignment by directly editing its weights (AlphAxiv)
2026-11-13: Visualizing thousands of LLM conversations in 3D (Operator0x)
2026-11-08: Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU (Github)
2026-11-06: Hope AI from Google retrains itself over “sleep”! (YouTube Short)
2026-11-05: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 (METR)
2026-11-03: I gave Claude ~700 ransomware encrypted files. It told me upfront it couldn’t decrypt them, then got two thirds of them usable again (Reddit). Here’s a more technical deep dive on Medium
2026-11-01: Resisting Cognitive Surrender (knlb.dev).
2026-02-28: Relevant: AI Isn’t Replacing SREs. It’s Deskilling Them from Elizabeth
2026-11-01: DeepSeek Open-Sources dsh-superagent, a Router-Worker Coding Agent (Github)
2026-10-31: Security issue found in the official Claude Code VS Code Extension (Phorion)
2026-10-31: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard, there are caveats though: trading wall clock for quality, low sample data, and verifying against known solutions (Github)
2026-10-30: I Read the Research on AI Hacking. Here’s What Actually Works, El analyses the increase in AI-driven security incidents in relation to the upcoming IPO (YouTube)
2026-10-30: Using AI to explore ideas that are unexplored and outside the training (YouTube)
2026-10-30: My own tests showing how prediction models (both LLMs and Decision Models) struggle with randomness (Github)
2026-10-29: Why AI Is Obsessed With Weird Sentences, BetterStack breaks down a research on how post training drifts the models (YouTube Short)
2026-10-28: I Built My Own Intelligent AI Gateway, Tim shows a cool multi-model local AI setup (YouTube 19m)
2026-10-28: The AI cost of JavaScript frameworks (Nolan Lawson)
2026-04-16: Removing Google SynthID using a phase-shift attack (YouTube Short) and here’s the repo.
If you found this issue interesting please leave a like ❤️ and share it to help its growth.


That was a very nice read! Looking forward for more
For the Strands Decider swap I'd also diff the label distribution week over week, alongside per-ticket shadow accuracy. When a new product line lands, a decision model keeps scoring fine on the old categories while quietly dumping the new ones into the nearest bucket.