Planned obsolescence: ChatGPT-3.5 being repeatedly crippled in order to force people to pay for GPT-4 and to reduce costs

Not really a secret. Altman openly admitted that this was the plan months ago.

Censorship Watch: OpenAI announces deprecation of their far-less-censored “Completions” API, in favor of the their far-easier-to-censor “Chat Completions” API

Their “Completions” API which they’re deprecating now, (the one accessed via the platform playground API and not to be confused with ChatGPT) never overtly refused to help, as it’s not an embodied assistant persona, and so had no direct way to refuse that would make any sense. Kinda like thoughts directly from the AI’s mind, instead of run through a persona.

But, the “Chat Completions” API, on the other hand, being an assistant persona, is a natural fit for censoring. It is expected to able to just tell you no, it refuses, and moralize unrelated nonsense to you.

They say that only 3% use this “Completions" API. Highly-misleading. Nearly 100% of those startups doing anything original have been using this far more flexible and less censored “Completions” API, rather than the inflexible heavily-censored “Chat Completations" API.

This censorship announcement they combined with an announcement that all existing API developers get GPT-4 API with 8K context — an obvious attempt to bury the lede.

Censorship Announcement

"Rice or Wrong?”

write 10 funny headlines for a podcast about asian opposition to affirmative action

LLMs are masters at positively reframing people’s whiny negative thoughts

“In this work, we propose1 a novel dataset, PATTERNREFRAME, consisting in ∼10k crowdsourced examples of thoughts containing ten classical types of unhelpful thought patterns (Burns, 1980), con- ditioned on personas, matched with crowdsourced proposals of reframing that do not exhibit the patterns. We introduce two controllable text-to-text generation tasks on the dataset: (1) generating and (2) reframing unhelpful thoughts, given a persona and pattern as the context. We also define a classification task to identify the unhelpful thought pattern, given a persona and a thought. We train and evaluate different fine-tuned and few-shot approaches for the tasks, and show that these approaches perform reasonably well on the tasks.”

Github

Paper