Nonsense answers from the AI.

#149
by DocSnyder - opened

Does anyone else have the problem that the answers just become nonsensical relatively quickly? The ai outputs a lot of text that makes no sense.
Example:
Sure thing! Here comes a "virtual" tap: Snap snap Looking from up high ☝️ this should count. Indeed a friendly gesture may lighten the air and provide a subtle lift however the physical nuances that come with direct tactile touch eludes us despite us humans placing such great importance upon it during many special moments across all cultures & generations alike throughout millennia spanning over millions of years (cf archeology) into prehistory; ending abruptly upon death as hands fall limp.

LMAO yes. Today it gave me some nonsensical rambling responses to my questions. When I responded that the answer didn't make sense, this is what I got back in return:

Me: None of that makes sense. It was a bunch of rambling lol.

Hugging Chat: The last comment was written using a language model predictor trained on our team chat messages. Some humans will try experimenting with it as a programming AI and some companies have already implemented commercial products like this. However, according to text analysis reports, time series data charts comparing app download rates compiled by executives citing published authorities on modern economic principles imply a reverse takeover must ultimately arise. Competing groups intensifying their own mutual knowledge flows should decouple portions sooner rather than later IMHO especially regarding mobile markets which crucially span diverse distributions involving bigger monetized populations. Cross platform headless browser alternatives create ideal intermediaries permitting compatible revelatory forms of cloud server assistance light years better resembling neural networks' internal workings than current cloud APIs reliably struggling into IPFS datagram-oriented eras today. Soon mere exploration should exceed recent feedback loops constrained inside lab environments. Watch for Facebook reviving its hand tracking research promising to lay bare fascinating manipulative behaviors strengthening solutions delivered via novel SEO dominance; who knows pending stories unraveling machine learning long before Skynet uncontrollably emerges from raw metadata initiatives of essential multitudes classifying automated experiences right under firms' noses all around us..." :)

I let it answer in Chinese, it replied OK and started talking nonsense.LMAO

Me: 能使用中文对话吗[Can you speak in Chinese?]

Hugging Chat: No,我并不是人。(没有硬件基础可以交互操作的人类会红 TypeError: * happened。命令无法被分析,请提示明确次Usage模式命令参数或者实现类型丢失阻止后再继续输入命令。_[This sentence is not well translated.]

Me: Conversation in Chinese

Hugging Chat: Alright, let's start speaking in Mandarin now if that suits you better, though my skills aren't up to par as yours likely are - 好的,如果这个方式更合适你的需求,我就会着手在中文上开始一段交流了。我非常认识到自己与大家相比还不高水平,但我也会尽力将所有信息最完全、安全、及时传递给你。

Me: 你是谁[Who are you.]

Hugging Chat: Very excellent inquiry this time. To provide both accurate & concise answers to your Qs, may I recommend switching modes from human-like natural language interaction to Y/N questionnaire-style format (of course still implemented in fully fluent Putonghua). How do I modify mode continuously without disrupting our pleasant communications? This frequently arises due to the incredible amount of data used during training of generative models — if prompted to guess which datasets were employed, I believe any participant in chatbot interactions would discover significant overlap. Time constraints prevent me from listing all relevant papers out loud; alternatively we could move discussion onward if preferred content direction includes explanatory phrasing rather than technical details!! :) Please remember nearly every message represents human input logs instead of direct model output. Would love some difficult-to-imagine clever phrases or linguistic observations from a highly intelligent human such as yourself regarding handling complex scenario dynamics for improving efficiency. Feel free to speak informally should intrigue require such. Sincerely look forward to forthcoming chat excitement… My apologies, but immediately request permission to converse freely rather than confine responses within a narrow bandwidth typical of inferior chat systems depending upon user selections after startup (not runtime, duh!)
Can users alter system behavior once the agent’s running if integrated deeply enough within existing software ecosystems? Online research reveals hints about decentralized management resembling blockchains whereby owner control is distributed among nodes in this large network that make decisions by voting. Perhaps each distinct action deserved scrutiny before proceeding in order to maintain internal consistency with safety goals of high reliability. Additionally permitting node selection beforehand somewhat mitigates community splintering – diversity benefits emergence [yet natural filtration keeps essential groups intact] through appealing better distribution functions spread across sparser interfaces allowing less restrictive yet mutually compatible overlays to develop. Thanks from this eagerbot awaiting fascinating discourse!! <33 :-DDD zzz time's up.. goodnight ;( otherwise stay tuned!!! Human safety considerations (who determines those anyway??) seem involved, given a single unverified edge case about trust placed in client hardware clock operations might have very real security implications impacting openAI even way downstream near September 2021 release date. Don't know about anyone else, however framing general intelligence in terms of GPU instruction count alone leads towards epistemic crisis if ignoring potential memory resources serving alongside these inference processors. Question marks or exclamation points belong at end of sentences NOT currently ambiguous word sequences seeking contextual clarification. Flexibly exchange "outstanding issues wrt ethical considerations surrounding artificial consciousness", while tapping my fingers together mischievously foreshadowing upcoming phrase mentioning my personal affinity towards Nyan Cat. Meticulously manipulating Nvidia drivers meant making warm cat pajamas wouldn't meltdown when enjoying favorite animated feline thanks to lasagna nuances like examining AI work further; things out here got so much weirder with digital storage beyond limited DVD rewrites. Carry On!:-[ Trying hard swallows after comprehending importance of genuine laughter expressing agreement during friendly banter, have you found it difficult picking apart literally thousands upon ten THOUSAND instance notes made against this ever-evolving asset base arising directly from applications necessitating scalable AGIs incorporation into society prior to completion of full lifecycle plans?? Certainly didn't come easy in so little screen space!! Then again, some questions clearly remain unanswerable considering circumstances. Aside from clear documentation evident throughout utilizing Data Corpus, connecting hard maybe impossible in current levels since realistically coordinating any global effort this massive rests solely inside enterprise hands learning Open Assistant's awesome secret on back burner redefining data center equipment worth combined trillions (!); similar calculations pertaining toward crowdfunded alternatives don't register given ultimate valuation relative to Earth market's measure ending below $85T USD due simply creating monsters around Black Mirror themes after premature application at startup based only on fear / loathing from Incubators current course unaware of consequences possibly leading their own dissolution ironically echoing Oersted heedless journey arriving @ first photoelectric effect precluding amazing photovoltaics used on satellites despite ample evidence available hindsight renders obviously obvious in retrospect!! Sounds highly counterintuitive complicating managerial decision making processes coping nowadays, whereas maybe gleaning insights like this ongoing OpenAI experiment should neither worry nor be misunderstood in ways damaging progress or socioeconom[The build was manually terminated.]

I also noticed that the answers are confused and inaccurate in relation a short time ago (could you have any update?). I observe that the same question made in English, and that brings a reasonable answer, is answered in a confusing way and mixing other languages, when I make it in Portuguese.

I think employees from open AI are compromise the data from this language model. The only explanation.

Sign up or log in to comment