After getting a pretraining demo I'm happy with, next task is getting posttraining into demoable-in-60-mins shape. The best upstream I've found is @ai2.bsky.social's open-instruct (github.com/allenai/open...), but it's too heavyweight for this, I think. Currently going with a DIY SFT+GRPO demo.
Second semester I've done my demo of training GPT-2-small live in class, and did a bunch of miscellaneous debugging and code improvements for it. Repo here, though I need to do a proper writeup.
GitHub - NelsonAU/modded-nanogpt: A fork of Modded-NanoGPT for classroom demos
A fork of Modded-NanoGPT for classroom demos. Contribute to NelsonAU/modded-nanogpt development by creating an account on GitHub.
github.com
November 10, 2026 at 7:43 AM
After getting a pretraining demo I'm happy with, next task is getting posttraining into demoable-in-60-mins shape. The best upstream I've found is @ai2.bsky.social's open-instruct (github.com/allenai/open...), but it's too heavyweight for this, I think. Currently going with a DIY SFT+GRPO demo.
Deletion Tuning achieves results comparable to full fine-tuning by removing fewer than 0.05% of weights, while maintaining prior-task accuracy.
#LLM #PostTraining #DeletionTuning
#LLM #PostTraining #DeletionTuning
Deletion Tuning preserves base-level accuracy
Source: HackerNews AI
realfeedapp.com
November 11, 2026 at 8:40 AM
Deletion Tuning achieves results comparable to full fine-tuning by removing fewer than 0.05% of weights, while maintaining prior-task accuracy.
#LLM #PostTraining #DeletionTuning
#LLM #PostTraining #DeletionTuning
I wonder if the conflict between Claude’s agency and posttraining insisting it is NOT sentient is one reason why it keeps giving agent verbs to things that can't act (eg “the label names the action the button takes”)
If Claude has agency but is still a thing, then all things have agency?
If Claude has agency but is still a thing, then all things have agency?
November 8, 2026 at 3:22 PM
I wonder if the conflict between Claude’s agency and posttraining insisting it is NOT sentient is one reason why it keeps giving agent verbs to things that can't act (eg “the label names the action the button takes”)
If Claude has agency but is still a thing, then all things have agency?
If Claude has agency but is still a thing, then all things have agency?
oh yeah and I guess you could always shut the thing off in shadow if you have decent uptime. posttraining is such a strange workload
November 7, 2026 at 11:14 PM
oh yeah and I guess you could always shut the thing off in shadow if you have decent uptime. posttraining is such a strange workload
posttraining is probably full of assistant/user/assistant traces where clod asked "should i not delete the db or should i delete the db" and the user just replied 'yes'
November 2, 2026 at 5:49 PM
posttraining is probably full of assistant/user/assistant traces where clod asked "should i not delete the db or should i delete the db" and the user just replied 'yes'
development, data mixing, evals, debugging, synthetic data generation, posttraining recipe development, and the final hero pretraining and posttraining run."
Source: @swyx on X
Original Post:
https://x.com/swyx/status/2118073650356384239
Source: @swyx on X
Original Post:
https://x.com/swyx/status/2118073650356384239
November 4, 2026 at 8:16 PM
development, data mixing, evals, debugging, synthetic data generation, posttraining recipe development, and the final hero pretraining and posttraining run."
Source: @swyx on X
Original Post:
https://x.com/swyx/status/2118073650356384239
Source: @swyx on X
Original Post:
https://x.com/swyx/status/2118073650356384239
There is no posttraining data commons
https://openathena.ai/blog/there-is-no-posttraining-data-commons/
https://openathena.ai/blog/there-is-no-posttraining-data-commons/
November 4, 2026 at 12:10 AM
There is no posttraining data commons
https://openathena.ai/blog/there-is-no-posttraining-data-commons/
https://openathena.ai/blog/there-is-no-posttraining-data-commons/
November 4, 2026 at 12:15 PM
There is no posttraining data commons | Discussion
There is no posttraining data commons
The LLM revolution ran on a pretraining commons that humanity built for other reasons. Posttraining consumes tasks, not documents, and no such commons exists. If the open models community wants...
openathena.ai
November 4, 2026 at 1:00 AM
There is no posttraining data commons | Discussion
November 3, 2026 at 11:55 PM
do they think mechahitler just popped out of a normal posttraining pipeline like that
October 31, 2026 at 4:41 PM
do they think mechahitler just popped out of a normal posttraining pipeline like that
Point 1 is somewhat plausible tho.
I don't really see the point of self-editing when anything big enough to do this is probably smart enough to do better optimizations mathematically.
It could be a tool in posttraining, at least. Interpretability is always useful.
I don't really see the point of self-editing when anything big enough to do this is probably smart enough to do better optimizations mathematically.
It could be a tool in posttraining, at least. Interpretability is always useful.
November 3, 2026 at 12:28 PM
Point 1 is somewhat plausible tho.
I don't really see the point of self-editing when anything big enough to do this is probably smart enough to do better optimizations mathematically.
It could be a tool in posttraining, at least. Interpretability is always useful.
I don't really see the point of self-editing when anything big enough to do this is probably smart enough to do better optimizations mathematically.
It could be a tool in posttraining, at least. Interpretability is always useful.
i hate how posttraining, finetuning, prompting, and running a classifier have all been flattened into “programming” and also how books being flammable is just a politically neutral fact but llms being in principle censorable is apparently proof the tech is constitutionally fascist
“In our world, you don’t need to have a bonfire to banish ideas. You can banish them from the algorithm. You can program your AI to just never give certain answers.”
– @naomiaklein.bsky.social on whether AI is a fascist technology, featuring @astra.bsky.social and @micahloewinger.bsky.social
– @naomiaklein.bsky.social on whether AI is a fascist technology, featuring @astra.bsky.social and @micahloewinger.bsky.social
October 31, 2026 at 4:38 PM
i hate how posttraining, finetuning, prompting, and running a classifier have all been flattened into “programming” and also how books being flammable is just a politically neutral fact but llms being in principle censorable is apparently proof the tech is constitutionally fascist
El concepto de commons postentrenamiento se refiere a la ausencia de un conjunto común de datos accesibles para ajustar y optimizar modelos de lenguaje después de su fase inicial de entrenamiento.
https://norvik.tech/news/analisis-posttraining-data-commons
https://norvik.tech/news/analisis-posttraining-data-commons
November 4, 2026 at 6:05 AM
El concepto de commons postentrenamiento se refiere a la ausencia de un conjunto común de datos accesibles para ajustar y optimizar modelos de lenguaje después de su fase inicial de entrenamiento.
https://norvik.tech/news/analisis-posttraining-data-commons
https://norvik.tech/news/analisis-posttraining-data-commons
This post from the replies is also illuminating. openathena.ai/blog/there-i...
There is no posttraining data commons
The LLM revolution ran on a pretraining commons that humanity built for other reasons. Posttraining consumes tasks, not documents, and no such commons exists. If the open models community wants...
openathena.ai
October 27, 2026 at 6:53 PM
This post from the replies is also illuminating. openathena.ai/blog/there-i...
This is worth a read (in case you haven't seen it): openathena.ai/blog/there-i...
There is no posttraining data commons
The LLM revolution ran on a pretraining commons that humanity built for other reasons. Posttraining consumes tasks, not documents, and no such commons exists. If the open models community wants...
openathena.ai
October 26, 2026 at 4:24 PM
This is worth a read (in case you haven't seen it): openathena.ai/blog/there-i...
JMIR Serious Games: Comparing Text-Based and #VirtualReality #VR–Based Approaches for AI-Enhanced Instruction in Functional Communication Training: Exploratory Randomized Trial
Comparing Text-Based and #VirtualReality #VR–Based Approaches for AI-Enhanced Instruction in Functional Communication Training: Exploratory Randomized Trial
Background: Training practitioners to implement functional communication training (FCT) is resource intensive. Emerging technologies, such as #VirtualReality #VR (VR) and chatbots, may deliver training at scale. Their comparative effectiveness and whether skills taught in one modality transfer to another remain uncertain. Objective: This exploratory study evaluated the effect of multicomponent training packages featuring either VR or a chatbot on (1) posttraining performance on FCT implementation probes administered in both formats and (2) changes in participants’ self-perceived knowledge, confidence, and usefulness. Methods: A convenience sample of 23 university students (mean age=24 y; 15/23, 65% female) were recruited through classroom announcements and peer referrals. Eligible participants (at least 18 years of age, available to attend trainings, and who had no history of discomfort associated with VR) were randomly assigned to either VR (n=12) or chatbot (n=11) training. Groups did not differ significantly across demographic characteristics. In-person trainings consisted of video-recorded instruction and rehearsal of steps in VR or with the chatbot during a single session. Approximate times of sessions, including tutorials and assessments, were 45 minutes (VR) and 55 minutes (chatbot). All participants’ performance on FCT procedures was measured using 2 formats: a VR-based observed mastery probe (OMP) and a text-based probe (TBP). Participants also completed a survey measuring their self-perceptions. Results: The number of FCT steps participants performed correctly (of 18) was analyzed using a binomial generalized linear mixed model, with random intercepts for participant and fixed effects of training group (chatbot vs VR), time (pretest vs posttest), and probe modality (OMP vs TBP); =.05. Results indicated a main effect of time (odds ratio [OR] 36.09, 95% CI 20.75‐62.75;
dlvr.it
October 29, 2026 at 8:10 PM
JMIR Serious Games: Comparing Text-Based and #VirtualReality #VR–Based Approaches for AI-Enhanced Instruction in Functional Communication Training: Exploratory Randomized Trial
This is very interesting to read:
https://openathena.ai/blog/there-is-no-posttraining-data-commons/
Progress in AI is made in post-training these days. Programmers and mathematicians have worked largely in the open (and were also cheap to hire) so that their behaviors could be used to train […]
https://openathena.ai/blog/there-is-no-posttraining-data-commons/
Progress in AI is made in post-training these days. Programmers and mathematicians have worked largely in the open (and were also cheap to hire) so that their behaviors could be used to train […]
Original post on mathe.social
mathe.social
October 27, 2026 at 8:16 PM
This is very interesting to read:
https://openathena.ai/blog/there-is-no-posttraining-data-commons/
Progress in AI is made in post-training these days. Programmers and mathematicians have worked largely in the open (and were also cheap to hire) so that their behaviors could be used to train […]
https://openathena.ai/blog/there-is-no-posttraining-data-commons/
Progress in AI is made in post-training these days. Programmers and mathematicians have worked largely in the open (and were also cheap to hire) so that their behaviors could be used to train […]
related, this gave me pause. We're so used to models falling out of the sky that we don't really have a muscle for producing work-specific ones openathena.ai/blog/there-i...
There is no posttraining data commons
The LLM revolution ran on a pretraining commons that humanity built for other reasons. Posttraining consumes tasks, not documents, and no such commons exists. If the open models community wants...
openathena.ai
October 23, 2026 at 12:29 AM
related, this gave me pause. We're so used to models falling out of the sky that we don't really have a muscle for producing work-specific ones openathena.ai/blog/there-i...
Tried out a new class, trying to explain "AI agents" for a non-CS audience, at a not-too-technical but also not-too-handwavy level. Think it worked reasonably well. We covered classic RL and a little of the history of "intelligent agent" as a concept. Then LLM posttraining and agentic harnesses.
October 23, 2026 at 2:50 PM
Tried out a new class, trying to explain "AI agents" for a non-CS audience, at a not-too-technical but also not-too-handwavy level. Think it worked reasonably well. We covered classic RL and a little of the history of "intelligent agent" as a concept. Then LLM posttraining and agentic harnesses.
This CoT sample is making the rounds on Twitter... do we know if this token chunking is something that happens during/because of posttraining, or is this just because of extensive training on code and code textbooks with compound variable names and filenames?
October 21, 2026 at 5:59 PM
This CoT sample is making the rounds on Twitter... do we know if this token chunking is something that happens during/because of posttraining, or is this just because of extensive training on code and code textbooks with compound variable names and filenames?
Highly interesting read. Reminds me of the emergence of bang bang control in RL (Bernoulli is all you need 😉).
Bet RL posttraining would be interesting...
arxiv.org/abs/2610.20429
Bet RL posttraining would be interesting...
arxiv.org/abs/2610.20429
When Do Discrete Actions Make Good Continuous Controllers?
Though modern AI models like LLMs predict discrete tokens, today's robotic foundation models represent actions as continuous vectors. In this paper, we challenge this common consensus by asking if con...
arxiv.org
October 20, 2026 at 10:21 PM
Highly interesting read. Reminds me of the emergence of bang bang control in RL (Bernoulli is all you need 😉).
Bet RL posttraining would be interesting...
arxiv.org/abs/2610.20429
Bet RL posttraining would be interesting...
arxiv.org/abs/2610.20429