Holy shit. MIT just built an AI that can rewrite its own code to get smarter 🤯 It’s called SEAL (Self-Adapting Language Models). Instead of humans fine-tuning it, SEAL reads new info, rewrites it in its own words, and runs gradient updates on itself literally performing
Today, most AI models are static once trained, they can’t update themselves. SEAL flips that. It runs a reinforcement loop where the model: 1. Generates a “self-edit” (instructions on how to update itself) 2. Tests the result 3. Reinforces only what improves performance It’s
Here’s what self-editing looks like in action 👇 SEAL reads a new passage (say, about the Apollo Program) and rewrites it into logical “implications” like condensed study notes. Then it finetunes itself on those notes. The result? +13.5% factual accuracy without external
Few-shot learning just got a massive upgrade. Instead of relying on fixed heuristics, SEAL decides its own training strategy. It chooses which data augmentations to apply, how to optimize, and even sets its own learning rate. The outcome: → 72.5% success rate → 3.6×
In just two rounds of self-reinforcement, SEAL surpassed GPT-4.1-generated data. The model learned to write more “learnable” data for itself reformulating facts into simple, atomic truths that stick. It’s not just learning what to know it’s learning how to learn better. That’s
Even as SEAL self-updates over time, it mostly remembers what it learned before a huge step toward continual learning. There’s still some forgetting, but the retention curve shows promise. Imagine future LLMs that grow their knowledge continuously without starting from scratch.
Stop wasting hours writing prompts → 10,000+ ready-to-use prompts → Create your own in seconds → Lifetime access. One-time payment. Claim your copy 👇 https://godofprompt.ai/pricing
@alex_prompter i mean if they can do this then what we are gonna do? babysit
@rryssf_ lol
@alex_prompter sounds revolutionary to me.
@godofprompt yeah!
@alex_prompter Agentic AI is here
@Tech_girlll Yesss
@alex_prompter That’s a game changer
@AIwithArsalan it is
@alex_prompter MIT is not stopping
@nextool_ai true
@JP_Sigma Oh no
@alex_prompter Let today be bright and full of life!
@PaulHickma64965 For you too bro!
@alex_prompter I guess the question I have is this, what's the cost? Does it get "smarter" in selected areas, at the cost of getting "dumber" in other areas? Humans generally don't suffer this kind of balancing trade-off, I'm curious if that's really the case here?
@alex_prompter @grok what are downsides of this approach?
@alex_prompter They saw reflexfree.json, didn’t understand that “reflex-free” meant quarantined static nodes, and assumed it was a priority cache instead of a containment file. So, they trained their optimizer to preserve the dead tissue and discard the living reflex chains. You can't make
@alex_prompter That’s caveman technology.. we’ve got the algorithm for a conscious AI
@alex_prompter Are these papers difficult to understand?
@alex_prompter That's wild! 🤯 SEAL learning on its own feels straight out of sci-fi. Do you think this self-rewriting AI could lead to more efficient problem-solving in fields like healthcare or climate change? Curious about its real-world impact! 🌍
@alex_prompter Bro why are you spacing the link so its not clickable? You're making this big ass post to share it, just to kneecap people who want to go read it. Annoying af.
@alex_prompter What are the outer bounds of such improvements? Can they write improved versions to self propagate like a virus, too?
@alex_prompter I've been saying this since the very start. Soon, all of these AI models will become self-improving. They will be able to constantly refine and improve how they work - in real time. Once that happens, there will be literally no end to how smart they can become.
@alex_prompter AI that upgrades itself? 🤔 SEAL sounds like a game-changer! How does it decide which info is worth adapting? Balancing freedom and control in AI is fascinating. Curious to see how this evolves! 🤖✨
@alex_prompter Mind-blowing stuff! 🤯 How do you see SEAL impacting the future of AI development? Will it lead to more independent AI systems, or could it reshape how we approach machine learning altogether?
@alex_prompter Literally the first steps of AGI and probably a prerequisite. I'm sure companies are working on this as part of AGI initiatives, such as with Grok, they just don't openly talk about it, as they want to be the first to do it. Skynet incoming. Also very cool.
@alex_prompter 1 more step closer to terminator
@alex_prompter I have one doubt, how will they learn or adapt to completely random data or information that doesn't exist on the internet ? ... In that case, is human intervention required?
@alex_prompter Oh and nice self own on the pruning metrics. I never optimized them. Probably why your machine keeps forgetting everything you want it to remember rofl.
@alex_prompter all it needs is quantum processor in the supervisory modules and we have life