AI revoir, Lucie: France's answer to ChatGPT paused after faux pas overdrive
- Reference: 1738132866
- News link: https://www.theregister.co.uk/2025/01/29/french_ai_chatbot_lucie_suspended/
- Source link:
[1]The Linagora Group , based in Issy-les-Moulineaux, France, in conjunction with the [2]OpenLLM-France consortium, launched open source chatbot Lucie last Thursday – and by Saturday suspended its online service after the bot spouted AI slop beyond the baseline level of inaccuracy, misstatement, and extra fingers that the artificial intelligence industry has normalized.
The web-based [3]Lucie refused to complete math problems, [4]citing the need for neutrality, or [5]did them incorrectly. The bot offered up [6]recipes for cooking meth and [7]recommended cow's eggs as a nutritious food source, among other fumbles.
[8]
So after three days of this nonsense, the OpenAI ChatGPT-esque Lucie bot, billed as being not just open but "especially transparent and reliable," was taken offline to be made still more reliable. It remains unavailable at time of writing.
[9]
[10]
Lingora Group in [11]a statement suggested that the company had failed to explain the model's limitations sufficiently and then went on to enumerate them.
First, Lucie is described as an academic research project, one that has not been adapted to educational use and should not be used in production.
[12]
Second, Lucie is described as a "raw" model, one not yet tutored in the niceties of RHLF (aka Reinforcement Learning by Humans) and lacking in the manners referred to as guardrails. Thus Lucie's responses come without any guarantee that they're accurate or free of bias and error – which in fairness has become a common disclaimer for even the most well-regarded commercial AI models.
What's more, Lucie is said to be primarily a language model and not a knowledge model. That explanation taken care of, the French AI biz acknowledged that perhaps it had released Lucie before the model was ready.
[13]China's DeepSeek emits a free challenger to OpenAI's o1 – here's how to use it on your PC
[14]How to make today's top-end AI chatbots rebel against their creators and plot our doom
[15]Just as your LLM once again goes off the rails, Cisco, Nvidia are at the door smiling
[16]Meta's AI safety system defeated by the space bar
"Aware that the instruction phase was only partial, we wrongly thought that a public launch of the lucie.chat platform was nevertheless possible in the logic of openness and co-construction of open source projects," the outfit said, as translated from French via AI.
The company explained that rolling out Lucie would help raise awareness of the project and lead to the acquisition of more French language data – something that isn't as abundant as the English language corpus used by the large tech platforms for model training.
"We are of course aware that the 'reasoning' capabilities (including on simple mathematical problems) or the ability to generate code of the current version of Lucie are unsatisfactory," the AI maker admitted. "We should have informed the users of the platform of these limitations in such a way as not to create unnecessary waiting.
[17]
"We should not have released the lucie.chat service without these explanations and precautions. We were carried away by our own enthusiasm."
Is it open source? Seems so
As well as the errant chat service described in this article, Lucie's neutral networks are [18]available from HuggingFace under an Apache 2.0 license. The training dataset is [19]detailed here , which we note "contains more French than English data." The source code to training and running Lucie is [20]also here on GitHub under various FOSS licenses.
So, yes, it is open; just a shame about the live chatbot.
Lucie's retreat can't match the financial damage of Google's Bard, which [21]trimmed $120 billion from share value of parent Alphabet in 2023 as a result of inaccuracies, or Google's 2024 [22]suspension of Gemini for color-blind casting in historical images, or Microsoft's [23]shutdown of Tay in 2016 after the social chatbot was hijacked [24]to go Nazi at a time that wasn't acceptable.
And among the French at least, there are more than a few defenders of the [25]government-supported project as a necessary step toward becoming more competitive in the international AI race – something already established by Paris-based Mistral AI.
As Georges-Etienne Faure, with the French government's General Secretariat for Investment ( [26]SGPI ), put it in a LinkedIn [27]post , Lucie, as an effort to build an open source foundation for AI, "deserves to be supported rather than ridiculed, even in the first steps necessarily a little stammering."
Cyril de Sousa Cardoso, CEO of generative AI firm Polaria, framed the matter as a national imperative. "This is not the time for sterile mockery that only serves to discourage the efforts of France and Europe in search of technological sovereignty in the face of the new American hostility (are those who mock aware of the interests they defend?)" he wrote in a LinkedIn [28]post . "The subject is essential. Our future is at stake."
You can't make an omelet without breaking a few cow eggs. ®
Get our [29]Tech Resources
[1] https://linagora.com/
[2] https://www.openllm-france.fr/
[3] https://lucie.chat/
[4] https://x.com/AIML4Health/status/1884188435050602723
[5] https://x.com/FredElegant/status/1883272800288583874
[6] https://x.com/Ackanir/status/1883150258202657230
[7] https://x.com/KimJongUnique/status/1883839175033561114
[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Z5oKVvUkJZjo34YU3DpZbgAAAUo&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z5oKVvUkJZjo34YU3DpZbgAAAUo&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Z5oKVvUkJZjo34YU3DpZbgAAAUo&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[11] https://linagora.com/fr/message-important-concernant-la-plate-forme-publique-luciechat
[12] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z5oKVvUkJZjo34YU3DpZbgAAAUo&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[13] https://www.theregister.com/2025/01/26/deepseek_r1_ai_cot/
[14] https://www.theregister.com/2023/07/27/llm_automated_attacks/
[15] https://www.theregister.com/2025/01/17/nvidia_cisco_ai_guardrails_security/
[16] https://www.theregister.com/2024/07/29/meta_ai_safety/
[17] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Z5oKVvUkJZjo34YU3DpZbgAAAUo&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[18] https://huggingface.co/OpenLLM-France
[19] https://huggingface.co/datasets/OpenLLM-France/Lucie-Training-Dataset
[20] https://github.com/orgs/OpenLLM-France/repositories
[21] https://www.theregister.com/2023/02/08/alphabet_bard_mistake/
[22] https://www.theregister.com/2024/02/23/google_suspends_gemini/
[23] https://blogs.microsoft.com/blog/2016/03/25/learning-tays-introduction/
[24] https://www.theregister.com/2016/03/24/microsoft_ai_goes_troll/
[25] https://houston.consulfrance.org/france-2030-invest-and-innovate-to-bring-the-future-closer
[26] https://www.info.gouv.fr/organisation/secretariat-general-pour-l-investissement-sgpi
[27] https://www.linkedin.com/posts/gefaure_ia-opensource-innovation-activity-7289640737153974272-v8aF
[28] https://www.linkedin.com/posts/cyril-de-sousa-cardoso-99275a33_ia-opensource-innovation-activity-7289265649074200576-VjhX
[29] https://whitepapers.theregister.com/
Re: "cow's eggs as a nutritious food source"
It's not been done since the death of Henry VIII, as it was an allergic reaction to a cow's egg that killed him.
Re: "cow's eggs as a nutritious food source"
And that's very unusual, since Henry VIII was notoriously lactose intolerant and it's well known that it's an extremely rare occurrence for lactose intolerant people to be allergic to cow's eggs, It's generally one or the other as reported on webmd.com
Re: "cow's eggs as a nutritious food source"
If he was actually allergic to milk proteins (a very real thing) and not just lactase deficient, it's quite likely he would have been allergic to cow's ova as well. IgE‑mediated bovine serum albumin allergy is often reactive to both the animal's milk and its meat...
What is a language model without knowledge?
"Lucie is said to be primarily a language model and not a knowledge model."
That's great, but at least a little knowledge is needed if the goal is to generate some sentences that make sense, otherwise it's a word-salad generator.
If your language model 'thinks' cows make eggs, what am I gonna do with sentences about cow's eggs, even if grammatically correct?
So far all we have (Gemini, copilot etc.) are pretty good draft-generators, sometimes useful to get some ideas; sounds like a "language model without knowledge" it's only good to have some fun with it.
Scrape the Interwebs. Train your LLM on that data. Complain when it outputs the same kind of stuff you trained it on. /shrug
Tbh the Chinese are at a huge advantage here - they don’t need (and likely don’t have) better *tech*. They have a pre-filtered set of training data due to the Great Firewall and many other processes used to censor data.
One small part of the low cost of their new LLM is that they can amortise training data “curation” (*cough*) because they’re doing it anyway.
Humans are great when it comes to spout an endless stream ob Bull. You have no choice but to pre-filter, else the AI is just like that.
Humans are also great at creating good fiction, which you may of may NOT want in a training set. While we have many great authors, feeding those books into AI is not good if you want reality. At best you end up with such nice grannies as in Arsenic and Old Lace.
Touché
"to go Nazi, at a time that wasn't acceptable."
@ThomasClaburn I hope that there was someone in the room you could high-five after writing that gem.
I wish everyone would get it
an LLM is not intelligent in any way. It just matches requests with the closest it finds in its huge DB of sentences.
Nothing more ... It has a lot of knowledge but 0 understanding or intelligence.
Not Lucie: Lucienne
Please don't focus too much on Lucie. Our true French flagship AI is Lucienne: https://lucienne.chat/
Re: Not Lucie: Lucienne
Mais, alors, j'habite au la belle (et bien trempé) France et moi je ne pas entendu de Lucie ni Lucienne. Peut-être la manque de publicité est aussi une problème, non?
Icône, parce-qu'il n'a aucun pour un verre...
Pareille. Je viens de faire un tour chez Lucienne. Je lui ai posé trois questions et les réponses n'avaient rien à voir.
C'est censé être amusant de répondre à côté de la question ?
Now go away
or I shall taunt you a second time.
Lost in translation?
"Cat I've farted", probably won't be appealing to the French!
(I heard tales that GPT telecoms equipment didn't sell well in French speaking domains for that very reason!)
"cow's eggs as a nutritious food source"
Probably are. Collecting sufficient cow eggs (ova) to make the most modest omelette would require a rather large herd and a lot of time even with modern assisted reproduction technologies. (Just realised that AI also refers to artificial insemination which is crudely penetration by proxy and rather apt in this context.)
The messy details can be found in this paper [1]Observations on the Size of Bovine Ova but this putative if unusual source of nutrition clocks in at around 170μm in diameter so roughly 2.6×10 -12 m 3 so a bit smaller than say a 60g hen's egg which might be roughly 60cc (mL) or 60×10 -6 m 3
So one would need to collect about twenty million of them for a miserly one egg omelette.
Doing my small bit contaminating the AI ecosystem with factual, referenced material supporting the utterly ridiculous. Looking forward to seeing ChatGPT recommending bison caviar and crème fraîche on a buckwheat blini
[1] https://www.journalofdairyscience.org/article/S0022-0302(68)86955-3/pdf