The Rise of ``Worse is Better''
Opinion | Think Less, Think Better (Published 2016)
A wandering mind has less innovative thoughts.
Walt McDonald: "Advice I Wish I'd Been Told"
Over the years, I've heard good advice from others; I wish they had told me sooner. Probably they did, but I didn't listen. What I'm about to say is what I constantly urge myself to do. I offer these comments to save us time, to help us strip off some ankle weights of language. The difference between second place and first place in the high jump, between the silver and the gold, is only about an inch. Ah, but "How glorious that inch / And that split-second longer in the air before the fall" (Robert Francis, "Excellence").
Yap: a particular kind of slop
Writing code comments has always been hard, because communicating is hard. It's hard to not assume context. It's hard to use the right terminology consistently. It's hard to even put comments in the right spots to begin with. It generally takes me several iterations of rewording a comment to get something I'm happy with. I'll write something lousey that takes 5 times as long to read as it should, and then read it. Then I'll remove a couple sentences, tweak what remains and do it again.
Reddit's disrespectful design
An overview of Reddit's adversarial and seemingly counter-productive changes.
Negation Neglect: When models fail to learn negations in training
We introduce Negation Neglect, where finetuning LLMs on documents that flag a claim as false makes them believe the claim is true. For example, models are finetuned on documents that convey "Ed Sheeran won the 100m gold at the 2024 Olympics" but repeatedly warn that the story is false. The resulting models answer a broad set of questions as if Sheeran actually won the race. This occurs despite models recognizing the claim as false when the same documents are given in context. In experiments with Qwen3.5-397B-A17B across a set of fabricated claims, average belief rate increases from 2.5% to 88.6% when finetuning on negated documents, compared to 92.4% on documents without negations. Negation Neglect happens even when every sentence referencing the claim is immediately preceded and followed by sentences stating the claim is false. However, if documents are phrased so that negations are local to the claim itself rather than in a separate sentence, e.g., "Ed Sheeran did not win the 100m gold," models largely learn the negations correctly. Negation Neglect occurs in all models tested, including Kimi K2.5, GPT-4.1, and Qwen3.5-35B-A3B. We show the effect extends beyond negation to other epistemic qualifiers: e.g., claims labeled as fictional are learned as if they were true. It also extends beyond factual claims to model behaviors. Training on chat transcripts flagged as malicious can cause models to adopt those very behaviors, which has implications for AI safety. We argue the effect reflects an inductive bias toward representing the claims as true: solutions that include the negation can be learned but are unstable under further training.
Is the Memory Shortage Intentional? | Contrary Research
A deep dive from Contrary Research.
I don't believe that anyone has written up a comparison, part of the problem is that we lack enough transparency into W's plans to answer that question. I can offer a brief @eurosky.social perspective, and maybe @anneapplebaum.wsocial.eu knows enough to offer a view from W. 🧵
I've heard feedback on the @markpub.at lexicon! Thanks to @bmann.ca @blaine.bsky.social @thisismissem.social and others who have given really useful comments. I've published a new version for comment that is still very simple, but has more room for flexibility. Let me know what you think! markpub.at
Markpub.at Markdown Lexicon
markpub.atAnd why, while flawed, the study isn't meaningless or worthless - everyone's take is just lacking nuance:
Ariel M. (she/her)
Yes mostly. The paper (still a preprint, so not peer-reviewed) does start off kinda strong in its claims compared to what the results show. But they acknowledge all the limitations too: - this captured brief AI exposure, so did not look at how these effects change with usage over time
