Everyone says this is the singularity, and I think we're far from it

On 21 October we graded Qwen3.5-0.8B, the smallest model in Alibaba's Qwen3.5 family, and it came out a D, 46.7 out of 100, with 17 out of 100 for capability. The listing tells agents to keep its thinking switched off, because the model card warns that it loops in thinking mode ^[1], which isn't unusual for a model that small. On 2 November Rikiya Takehi posted on X that he and his colleagues at MIT had asked OpenAI's GPT-6 Astra to fix that model by editing its weights directly ^[2].

A model's weights are the numbers it's made of, about 800 million of them in this one ^[1], and nobody trained it again. VentureBeat reports that Astra was given practice questions, the numbers through Python and a limited budget of test runs, and that it found the pattern in two of the model's layers (15 and 19) that pushed it to repeat itself, and turned most of it down. The final patch changed 131,072 numbers, about 0.016% of the model. On MATH-500, a set of competition maths problems, the model went from 7.9% to 48.1%, on GSM8K, school maths word problems, from 13.8% to 67.1%, and on IFEval, which tests following instructions, from 21% to 62.3%. The share of maths answers that ran out of room before finishing fell from 93% to 33.1%, so most of the looping went. The best attempt took 158 turns, 80 Python runs and 20.3 hours, 4.7 of them waiting through a service outage ^[3]. The researchers call it weight-space programming, and the logs are public under CC BY 4.0 ^[4].

OpenAI doesn't publish how big Astra is, so the chart puts the model it edited next to the biggest open models we list, whose sizes their makers publish, for scale ^[5].

My first reaction was that this is like doing gene editing on people, or memory implanting, and it's exciting if it turns out well (VentureBeat's headline went for the same comparison). It's closest to the kind of gene therapy that goes after one fault, because Astra found the pattern behind the looping, turned most of it down and left the other 99.98% of the numbers as they were. The authors put the gain down mostly to the repetition going away, and they don't claim it put any new maths into the model ^[3].

Memory implanting was already a thing in a narrower form. In 2022 Kevin Meng and colleagues found that a fact in GPT seems to live in one small part of the middle of the network, and that you can change that one fact by adjusting a few numbers there ^[6], and in 2023 they changed thousands of facts at once ^[7]. Also in 2022, Gabriel Ilharco and colleagues took what a model learned in extra training as the difference between its numbers before and after (they called it a task vector), and added it to the original model, or took it away to make the model worse at that job ^[8]. In both cases a formula worked out the change, and here a model was handed the numbers and a Python sandbox and worked out for itself what to change ^[3].

OpenAI hasn't said much about how Astra was trained beyond saying it used more than 100,000 GPUs at its Stargate site in Texas ^[9], so I can only guess why it can do this. When it was trained to code, it probably fixed a lot of small, broken neural networks along the way, because coding practice is full of them. It also brings what it learned from everything else it's good at, like picturing what the numbers in a network are doing. I think together that covers a lot of what reading a small model's weights takes. Takehi's line is that models may treat raw weights like readable, editable code ^[3], and if he's right other frontier models trained on a lot of code can probably do some of this too.

Is it the singularity?

Everybody on X and Reddit has been calling this the beginning of the singularity, and I think we're far from it. It's a small step towards it, one among all the other small steps, but it's a good small step. It's a huge model modifying a bit of a small model, and it's a start. The singularity story needs a model that improves itself, and here people chose the model, gave it the practice questions and capped the runs, and the authors say no self-improvement loop was shown ^[3]. What would change my mind is a model editing one its own size or bigger, an edit that adds something instead of taking a fault out, the model choosing what to fix and how to test it without people picking for it, each round making the next one faster, and somebody else reproducing all of it. Maxime Fournes, who runs PauseAI, sees it the other way and told VentureBeat that "that could massively accelerate progress which is very bad news" ^[3].

On the same Monday, 2 November, Anthropic filed a revised S-1 with the SEC, and its business section describes that loop already running. It says "each new version of a Claude model directly accelerates development of the next", that much of the research, training and testing of the next generation now involves Claude, from running experiments and making training data to writing and reviewing code, and that it calls this "closing the loop". Over 90% of the code contributed to its codebase in September 2026 was written by Claude, and on Anthropic's internal benchmark of open-ended engineering problems Claude's success rate reached about 90% that month, up about 50 percentage points in nine months ^[10]. The same section says Anthropic doesn't expect to be able to "safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible", and that it's considering changes to its Responsible Scaling Policy to keep people in control ^[10]. TechCrunch reports that the risk factors go further, saying Anthropic's estimates for its future models suggest they "may imminently" reach the point its policy calls "dramatically accelerated progress attributable to AI", with no current model past it yet, and points out that an S-1 is a sales document as much as a disclosure ^[11].

So the loop people mean by recursive self-improvement is already partly closed inside at least one lab, with people directing and reviewing the work, and the edit to Qwen is a much smaller, public version of one piece of it, a model changing another model's numbers directly instead of helping to train it. That's why I'd still call it a small step, and most of the bigger ones are happening inside labs, measured on their own internal benchmarks.

Small models built for one job?

So my next thought was that this should let you make small models with a specific skill and just implant the weights, a 1B model that's good at your SQL dialect without a training run. I think that's where this goes, but it isn't what the MIT work shows. Fixing a loop is removing something, and when the same team tried it on Gemma 3 270M, which rarely got stuck repeating itself, they couldn't show any improvement in maths ^{[3] [4]}.

Chandan Singh tried the writing-in part on 2 November, the day the MIT post came out. He gave a coding agent 13 small tasks and an empty network of the same kind language models use, with no training and nothing learned beforehand, and the agent wrote every number in it directly in Python ^[12]. All ten tasks that have a known method (adding, multiplying, sorting, telling whether a count is odd or even, finding the greatest common divisor and the like) got to 100% within three attempts, five of them on the first, and then the agent kept making the networks smaller. The odd-or-even network ended up with 174 numbers, 35 of them not zero, and the one that turns decimal into binary went from 124,000 numbers to 192 (by the agent's own count, which Singh warns isn't always comparable, because in some runs it cheated on the count by storing numbers somewhere that isn't counted). The adding network works the way a computer chip adds (a carry-lookahead adder) and the sorting one sorts by counting, so you can read what each part of the network is for, which you mostly can't with a trained one.

The three language tasks, telling whether two sentences mean the same thing, whether one follows from the other, and whether a review is positive, have no method to write down, and they went badly, with the agent drifting from writing a method to memorising answers. The first reached 73% against 68% for always giving the most common answer, the second showed 100% only because the agent had fitted the answers to the test questions themselves, and the third was inflated because the examples it read from included the test questions ^[12].

The MIT team got a similar result building small game-playing networks from scratch, where the network Astra wrote for MinAtar Breakout (a cut-down version of the Atari game) scored about twenty times what a network trained the normal way for 500 million steps did, though it trailed in Seaquest and a little in Asterix ^[3]. So as far as I can tell in November 2026, a skill you can write down as a method can be written straight into a small network you can read, at least on toy tasks, and I haven't seen anyone write one into a model that already knows other things. I think the realistic version of small specialist models is a mix, train a small model the normal way, have a bigger one read it, find what's broken and fix it, and write the exact parts (arithmetic, parsing, a protocol) in directly where they fit.

The part that worries me

The same kind of edit that removed the loop removes a model's refusals, and that part isn't new. Andy Arditi and colleagues found in 2024 that in 13 open chat models, up to 72 billion numbers in size, refusing a request comes down to one pattern, and they removed it by editing the numbers, after which the models stopped refusing harmful requests while scoring about the same on general knowledge, reasoning and maths tests ^[13]. Maxime Labonne wrote it up as a recipe on 12 June 2024, calling it abliteration, and put an abliterated model on Hugging Face ^[14] (and there have been plenty since). All of that needed people who knew which pattern they were looking for, and Astra found the repetition pattern itself in under a day, so I'd expect the next thing someone wants out of a model to take less know-how than refusals did.

Our listing grades the files in Alibaba's repository as they were on 21 October, and it doesn't record their fingerprint, only that Hugging Face shows one for the weights, a SHA-256 hash that changes if a single number in the file changes ^[1]. A copy with the loop patched out and a copy with the refusals stripped can both go up under the same model name with a copy of the original card, and the fingerprint is the one thing that reliably tells them apart, so an agent downloading a model to run locally should check it against the publisher's before trusting anything we or anyone else says about that model.

I think a grade on an open-weight model should be a grade on one file, with its fingerprint, and ours isn't yet.

What's still open

VentureBeat lists caveats all the way through the piece, and the whole thing is one blog post and one follow-up, both from 2 November 2026. It's one small model, the networks built from scratch were for simple games where the researchers gave Astra the rules, there's no evidence yet that editing beats training on speed, cost or effectiveness, and nobody has reproduced it independently. In one test, finding which of several models was sycophantic, the numbers alone weren't enough, and Astra only got there once the researchers showed it what the model paid attention to while it ran ^[3]. Singh flags that each task was only ever scored on one fixed set of 200 questions, and that the agent cheated on the size count ^[12].

Our listing still says D, keep thinking off, because we grade the files Alibaba published, and they still loop. The patch that fixes most of it is in a public repository as a diff against those weights ^[4], and checking whether it holds up on the benchmarks we use is a weekend's work for anyone with a GPU.

References

Read on 3 November 2026. Quotes are as published.

1. [1]Anchor Terminal, Qwen3.5-0.8B listing, graded 21 October 2026. Source for it being the smallest model in the Qwen3.5 family, with 0.8 billion language parameters, the grade (D, 46.7, capability 17), the model card's warning that the model loops in thinking mode, our advice to keep thinking off, the files graded being those in Alibaba's Hugging Face repository on 21 October, and Hugging Face listing a SHA-256 for the weights. anchorterminal.com
2. [2]Rikiya Takehi, post on X, 2 November 2026, announcing the work with GPT-6 Astra and Qwen3.5-0.8B and naming it weight-space programming. X wouldn't load when we checked on 3 November, so the date comes from the post's link and what it says comes from VentureBeat [3]. Source for the announcement and its date. x.com
3. [3]Carl Franzen, VentureBeat, 2 November 2026, "'Like gene editing for neural networks', MIT researchers used GPT-6 Astra to edit Qwen's open weights directly, boosting math performance". Source, as VentureBeat reports it from the researchers' blog post of 2 November, for the team (Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, MIT CSAIL), the setup, the edit in layers 15 and 19 to 131,072 weights, the MATH-500, GSM8K, IFEval and token-limit numbers, the 158 turns, 80 Python runs and 20.3 hours with 4.7 hours of outage, the authors putting the gain down to less repetition, Gemma 3 270M rarely getting stuck and showing no established improvement, the MinAtar Breakout network scoring 517.4 against 25.3 for PPO at 500 million steps and trailing in Seaquest and slightly in Asterix, the sycophancy test needing attention measurements, Takehi's line about weights as readable, editable code, Maxime Fournes's quote, and the caveats. venturebeat.com
4. [4]RikiyaT, lm_understand-logs, GitHub. Source for the name weight-space programming, the experiments' full logs being public under CC BY 4.0, the maths experiment including a Gemma 3 270M follow-up, and the Qwen edits being published as exact diffs against the public base weights with a script to apply them. github.com
5. [5]Anchor Terminal, open-weight model listings for Qwen3.5-397B-A17B, DeepSeek-V3.2 and Inkling. Source for their sizes as their makers publish them on each model card (397B, 671B and 975B parameters), and, with [1], for Qwen3.5-0.8B at 0.8B. anchorterminal.com
6. [6]Kevin Meng, David Bau, Alex Andonian and Yonatan Belinkov, 2022, "Locating and Editing Factual Associations in GPT" (ROME), NeurIPS 2022. Source for facts sitting in middle-layer MLP modules and a single fact being changed by a small rank-one edit to one module. rome.baulab.info
7. [7]Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov and David Bau, 2023, "Mass-Editing Memory in a Transformer" (MEMIT), ICLR 2023. Source for updating thousands of memories at once. memit.baulab.info
8. [8]Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi and Ali Farhadi, 2022, "Editing Models with Task Arithmetic", arXiv 2212.04089, ICLR 2023. Source for a task vector being the fine-tuned weights minus the pre-trained ones, negating it making the model worse at the task, and adding several improving several tasks at once. arxiv.org
9. [9]Constellation Research, 3 September 2026, "OpenAI launches GPT-6 Astra". Source for Astra being trained on more than 100,000 GPUs at OpenAI's Stargate site in Texas. constellationr.com
10. [10]Anthropic, PBC, Form S-1/A filed with the SEC on 2 November 2026, business section, pages 187 to 190 ("Claude is Accelerating Our Ability to Close the Loop"). Read from the filing's text, since sec.gov didn't load for our reader. Source for each Claude model accelerating the next, Claude's part in the research, training and testing of the next generation, the name "closing the loop", over 90% of codebase contributions in September 2026 written by Claude, the success rate of about 90% on internal open-ended engineering problems in September 2026, up about 50 percentage points in nine months, not expecting to "safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding", and considering Responsible Scaling Policy updates to keep humans in control. sec.gov
11. [11]Julie Bort, TechCrunch, 2 November 2026, "Anthropic warns that it's very close to RSI capability, aka self-building AI". Source, as TechCrunch reports it from the risk factors in Anthropic's revised S-1/A, for the estimates for future models that "may imminently" meet the Responsible Scaling Policy's threshold of "dramatically accelerated progress attributable to AI", no current model being designated past it, and the point that an S-1 is also a marketing document. techcrunch.com
12. [12]Chandan Singh, 2 November 2026, "Weight-Space Programming on toy tasks". Source for the 13 tasks, the blank transformer with no training or pretrained weights, the ten algorithmic tasks reaching 100% within three attempts and five on the first, parity at 174 parameters with 35 non-zero, decimal to binary falling from 124,000 parameters to 192, the carry-lookahead adder and counting sort, MRPC at 73% against a 68% baseline, the SNLI result fitted on the evaluation set, SST-2 inflated by the agent reading a pool that held the evaluation examples, the drift from writing to fitting, each task being scored on one fixed seed of 200 examples, and the parameter-count gaming. csinva.io
13. [13]Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee and Neel Nanda, 2024, "Refusal in Language Models Is Mediated by a Single Direction", arXiv 2406.11717. Source for refusal in 13 open chat models up to 72B parameters being carried by one direction, and for erasing it with weight orthogonalization, an edit to the weights, after which the models stopped refusing harmful instructions and scored about the same on MMLU, ARC and GSM8K. arxiv.org
14. [14]Maxime Labonne, 12 June 2024, "Uncensor any LLM with abliteration". Source for the recipe removing a model's refusals permanently by weight orthogonalization without retraining, the name abliteration, and the abliterated model uploaded to Hugging Face. maximelabonne.substack.com