Martin Signoux
@MartinSignoux
Followers
2,611
Following
1,334
Media
196
Statuses
2,226
AI policy @OpenAI / Tweeting in my own capacity
Paris
Joined June 2016
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Venezuela
• 2174168 Tweets
Rayssa
• 939138 Tweets
Lewis
• 241049 Tweets
Russell
• 184000 Tweets
Japão
• 182149 Tweets
Rebeca
• 174322 Tweets
Hamilton
• 167760 Tweets
Victoria
• 167641 Tweets
Lebron
• 155602 Tweets
Simone Biles
• 145112 Tweets
#BBNaija
• 103325 Tweets
Tatum
• 84958 Tweets
Kevin Durant
• 71031 Tweets
Serbia
• 67606 Tweets
Omar
• 63196 Tweets
julia soares
• 59550 Tweets
Léon Marchand
• 59203 Tweets
Embiid
• 57020 Tweets
スケボー
• 46596 Tweets
Gimnasia
• 40232 Tweets
#النصر_بورتو
• 34780 Tweets
Kerr
• 32403 Tweets
Cubs
• 31182 Tweets
flavia
• 24935 Tweets
Paredes
• 23246 Tweets
Alexia
• 21777 Tweets
Raça Negra
• 19212 Tweets
İsrail
• 18972 Tweets
Samantha
• 18121 Tweets
Ronnie
• 17709 Tweets
Flavinha
• 15509 Tweets
Zambia
• 14466 Tweets
Andy Murray
• 14147 Tweets
Shaun
• 13007 Tweets
Morel
• 12966 Tweets
フェンシング
• 12453 Tweets
Biggie
• 11782 Tweets
محمد السيد
• 11130 Tweets
Jade Barbosa
• 10820 Tweets
Pinned Tweet
🌍Glad to share that I’m joining
@OpenAI
’s Global affairs team as AI policy lead for Europe 🇪🇺
The months ahead will be fast-paced and important for the future of AI, so I couldn’t be more excited to join this team to advance AI policy alongside brilliant people.
1/3
14
6
124
What a year it's been for AI.
Anticipating what's next sounds very perilous, but I'll give it a try.
Here are eight AI predictions for 2024🔮
1/n🧵
👇
24
118
393
Kyutai is taking over the original mission of OpenAI.
The very same day
@kyutai_labs
emerges at the leading non-profit AI lab committed to open source,
@OpenAI
might very well begin its fall.
What a symbol for open science !
What an omen for France !
7
41
220
Evaluation is all we need.
We can’t keep talking about AGI and ASI in the abstract.
We need a taxonomy of Machine Intelligence (what are we talking about?) and better evaluation tools (how de we measure this?).
A 🧵
1/n
👇
15
25
126
III) So long LLMs, hello LMMs
Large Multimodal Models (LMMs) will keep emerging and oust LLMs in the debate; multimodal evaluation, multimodal safety, multimodal this, multimodal that.
Plus, LMMs are a stepping stone towards truly general AI-assistant.
5/n
4
14
136
As always,
@benedictevans
delivers a thought-provoking vision.
This year questioning the real impact of GenAI and LLMs...
1/2
2
6
113
Technology maturity isn’t sufficient to drive adoption
Insightful analysis of the rise of tractors in US farms, put in parallel with Generative AI, by
@TheEconomist
AI adoption and subsequent socio-economic impact will take years, if not decades
Time to prepare, not to panic
9
24
104
V) Small is beautiful
Small Language Models (SLMs) are already a thing, but cost-efficiency and sustainability considerations will accelerate this trend.
Quantization will also greatly improve, driving a major wave of on-device integration for consumer services.
8/n
3
11
101
👋 Welcome Kyutai, a new non-profit AI lab dedicated to open science with a stellar team and advisory committee 🇫🇷
3
7
86
We’re ending 2023 with only 13% left between Mixtral and GPT-4 on MMLU.
But most importantly, open models are here to stay and drive progress, everybody realised that. They will coexist with proprietary ones, no matter what OS detractors do.
10/n
1
3
86
VI) An open model beats GPT-4, yet the open vs closed debate progressively fades
Looking back at the dynamism and progress made by the open source community over the past 12 months, it’s obvious that open models will soon close the performance gap.
9/n
5
4
84
I) AI smart glasses become a thing 😎
As multimodality rises, leading AI companies will double down on AI-first wearable devices.
And what’s better than the glasses form factor to host an AI-assistant ?
2/n
5
7
74
...iterative improvements will make them “good enough” for various tasks.
Improvements in RAG, data curation, better fine-tuning, quantization, etc, will make LLMs robust/useful enough for many use-cases, driving adoption in various services across industries.
7/n
1
3
62
📚Un cartographie instructive des différents textes 🇪🇺 qui régulent le secteur du numérique
Sur les 104 textes référencés:
👉 77 ont été adoptés (ou en cours d’adoption) depuis 2019
👉 27 l’ont été entre 1987 et 2019
1
34
65
II) ChatGPT won't be to AI assistant what Google is to search
2023 started with ChatGPT taking all the light and ends with Bard, Claude, Llama, Mistral and thousands of derivatives
As commoditization continues, ChatGPT will fade as THE reference ➡️ valuation correction
4/n
4
10
63
IV) No significant breakthrough, but improvements on all fronts
New models won't bring real breakthrough (👋GPT5) and LLMs will remain intrinsically limited and prone to hallucinations. We won’t see any leap making them reliable enough to "solve basic AGI" in 2024
Yet...
6/n
1
2
62
VIII) Existential-risks won't be much discussed compared to existing risks
While X-risks made the headlines in 2023, the public debate will focus much more on present risks and controversies related to bias, fake news, users safety, elections integrity, etc
12/n
4
3
59
VII) Benchmarking remains a conundrum
No set of benchmarks, leaderboard or evaluation tools emerge as THE one-stop-shop for model evaluation.
Instead, we’ll see a flurry of improvements (like HELM recently) and new initiatives (like GAIA), especially on multimodality.
11/n
2
2
56
Advancing the detection of « voice deepfakes » through watermarking 👇
Meta presents Proactive Detection of Voice Cloning with Localized Watermarking
paper page:
In the rapidly evolving field of speech generative models, there is a pressing need to ensure audio authenticity against the risks of voice cloning. We present
2
56
219
0
6
50
🔍Interested in GenAI watermarking ?
Make sure to follow the work of
@pierrefdz
from
@Meta
&
@Inria
!
After 🖼️ image (Stable Signature) and 🔊 audio (Audioseal), he's now making strides in 📄 text-generation with
@RednasTom
(🐓 Bonus: it's made in 🇫🇷 )
3
14
52
There are also interesting benchmarks being developed.
1/ GAIA (General AI assistant) a benchmark by researchers from Meta (
@mialon_gregoire
@ylecun
@ThomasScialom
) and Hugging Face (
@clefourrier
@Thom_Wolf
)
15/n
2
6
47
Branches are close to the ears to deliver audio, cameras are close to the eyes to capture ego-centric input, plus they’re hands-free and comfortable
Meta is leading the way with Ray Ban, but think about the recent Open AI x Snapchat rumor 👀
We’re only getting started...
3/n
2
3
44
“With the appropriate guardrails, objective-driven AI will be intrinsically safe because they’ll be bound not to deviate from the objective we will set” says
@ylecun
5
7
43
👋 Morning from
@joinstationf
2nd and last day of pitch for selecting the start-ups of the
@Meta
@huggingface
@Scaleway
AI incubator program 🤗
Lot of talent and energy.
See you in January to discover the great things they’re building with open source AI 🚀
0
2
34
10 years of open and collaborative research
A decade advancing open science
Over a thousand open source projects; models, datasets or libraries
0
1
30
📗Yesterday, the French AI commission published its report and handed it to
@EmmanuelMacron
🎯It sets an ambitious, resolutely positive, yet realistically grounded, vision for this technology as a source of economic and social progress.
🇫🇷🇪🇺Il n'y a plus qu'à 🚀
5
6
32
After 3 amazing years, last week marked the end of my journey at
@Meta
. I’m beyond grateful for everything l’ve learnt and all the brilliant minds I’ve met. You are too many to name here, but here’s a heartfelt thank you!
A new chapter in AI policy is about to begin. Stay tuned!
1
0
29
#GAFA
are not the biggest threat to privacy,
#data
brokers are. In
#ChaosMonkeys
@antoniogm
shed light on this issue, and so does the
@FT
today.
1
8
27
Reading of the full presentation is recommended👇
2
1
26
Today
@banquedefrance
says it successfully tested wholesale
#CBDC
and that other tests will be conducted in the coming weeks.
That's still a test, but France is definitely paving the way for the
#DigitalEuro
cc
@BlockchainforEU
0
8
24
🇫🇷 foreign minister publicly states that 🇫🇷 is preoccupied by the HK situation while the presidential sherpa privately says to the 🇨🇳 foreign minister that 🇫🇷 understands 🇨🇳
One country, two speeches ?! Looks more like 🇨🇳 trying to pressure 🇫🇷 through selected leak.
Macron’s foreign policy chief promised China that France would not “interfere” in Hong Kong affairs.
“France is fully aware of the sensitivity of issues related to Hong Kong,” Chinese state media quoted Emmanuel Bonne as saying.
Strategic autonomy
14
71
113
2
10
20
👏"Nous croyons en l'open source, nous devons accélérer sur l'open source" affirme
@EmmanuelMacron
au sujet de l'IA lors
#VivaTech
🇫🇷 L'occasion d'annoncer près de €650M d'investissement dans l'écosystème d'IA et €200M dans le métavers et les technologies immersives 🚀
0
9
20
Speaking of taxonomy, DeepMind (
@merrierm
@jaschasd
@triswarkentin
@AllanDafoe
@AleksandraFaust
@ShaneLegg
) elaborated one with their Levels of AGI paper :
14/n
2
1
19
0
5
18
🚀Beau moment collectif pour faire rayonner l’IA ouverte et collaborative à
@joinstationf
hier à l’occasion de l’annonce des lauréats de notre programme d’accélération dédié à l’IA open source lancé avec
@HuggingFace
et
@Scaleway
7
4
15
@fandoetlis
Et c’est John le Carré qui a payé les publicités sur X et les camions qui circulaient dans Paris et Bruxelles ? 🤭
🧐Ouvrons la bibliothèque publicitaire de X (ci-dessous le rapport pour Control AI sur les 3 derniers mois).
Rien n'est laissé au hasard ce qui indique clairement le professionnalisme et la volonté d'influencer le processus démocratique de l'AI Act 🇪🇺
A dérouler 🧵👇 1/n
1
3
12
1
4
16
Fascinating insights in the tech competitive landscape from
@CompetitionProf
during
@CERRE_ThinkTank
webinar.
Competition among big tech is a dynamic game with entangled monopoly and oligopoly 👇
1
3
16
🗺 La carte et le territoire par
@IGNFrance
x
@CNNum
!
On y trouve des réflexions condensées, mais riches, notamment sur le Métavers et notre rapport au réel, par
@jeancattan
et
@basdevant_a
Une boussole pour penser le futur 🧭
Merci
@agathe_bougon
pour la recommandation
0
4
15
Une enquête ne serait pas de refus pour faire toute la lumière sur cette campagne sulfureuse, soutenue par la mystérieuse organisation
@ai_ctrl
😉
@reesmarc
@GrablyR
@Heroceane
@delf1
@Jamal_ELH
5/5
2
0
14
The report is available here in FR with a synthesis in ENG
.
0
5
15
We need more people building better evaluation tools, rather than just optimising models for existing ones.
Evaluation isn’t as sexy as building models, but it’ll ultimately drive of progress in the field. Notably because people will then optimise for better benchmarks.
11/n
1
0
12
💯 L'éthique n'est pas la morale.
Il ne s'agit pas d'édicter le bien et le mal selon une approche manichéenne. Il s'agit d'éclairer, de contribuer à trouver le juste équilibre entre différentes valeurs.
👏
@mmitchell_ai
et l'équipe d'🤗
@GiadaPistilli
@YJernite
@IreneSolaiman
1
3
13
@sebkrier
💯 ! same goes with AI ethics, for e.g fairness and bias being a proxy for defining what justice and fairness should looks like, reigniting decades of fascinating debates about the philosophy of justice
The danger is thinking that the wheel has to be reinvented for each of these
1
1
9
🧐Ouvrons la bibliothèque publicitaire de X (ci-dessous le rapport pour Control AI sur les 3 derniers mois).
Rien n'est laissé au hasard ce qui indique clairement le professionnalisme et la volonté d'influencer le processus démocratique de l'AI Act 🇪🇺
A dérouler 🧵👇 1/n
1
3
12
#vendredilecture
#IA
Deux ouvrages pour qui veut approfondir la question de l’IA par-delà la technique, notre rapport aux technologies, et les enjeux politico-philosophiques sous-jacents.
Matière à penser pour les intelligences (non) artificiels !
👇🧵
2
2
11
I’ll still be based in Paris (with frequent trips to Brussels), so reach out if you’re visiting and want to discuss anything AI.
I’m looking forward to working with everybody involved in researching, developing, and distributing more capable and safer AI systems !
2/3
6
0
11
2/ WorldSense: a benchmark from Meta (Youssef Benchekroun,
@megdrv
Mark Ibrahim) to evaluate the cohesiveness of world models in LLMs
16/16
2
0
11
👏An excellent primer to the regulation of foundation models in the AI Act by
@ckutterer
from
@AI_Regulation
Some quick and high-level remarks in this 🧵
👇
1/
1
2
9
Full article
0
5
11
Takeaway: we need better benchmarks. We need to invest more in evaluation, so as to build a scientific consensus, and then makes this understandable to a non-tech audience. This is vital to advance AI science AND to build social acceptance, by demystifying what AI really can
10/n
3
0
10
That being said, there are a few recent and promising initiatives aimed at advancing these taxonomy and evaluations.
I’m starting a list of ressources and articles below👇
Suggestions and comments are welcome !
13/n
1
0
9
A huge part of the moral panic that drove the public conversation these past months are grounded in the lack of litterature defining and discussing so-called AGI, and metrics related to models performance, capabilities and limits.
4/n
1
0
9
🥽Le métavers et la XR vous intéresse ?
🗓️RDV le 19 octobre pour le lancement du livre blanc de la Fondation
@simplonco
pour des techno plus responsables
🎤J’aurai le plaisir d’intervenir avec Maud Clavier du CNXR et
@VincentGuigui
@Accenture
sous la modération de
@fbardeau
2
4
10
Thank you
@sGianella
& al. for your trust.
And now let’s make AI a source of progress for all 🤝
3/3
1
0
9
2/ Gemini’s performance and comparison to others models are inconsistently presented (w/ 0-shot, X-shot, variable-shot, CoT
@32
…) across 32 benchmarks that are not equally robust, fairly limited in terms of intelligence evaluation, and mostly obscure to a non-tech audience.
8/n
1
0
9
Part of the answers to these questions are: better benchmarks and taxonomy.
And it would not only benefit the AI field, it’d also help decision-makers and society grasp the implications, ultimately driving better policies, governance frameworks and social acceptance.
3/n
1
1
8
The failure to build a taxonomy and better evaluation would result in a lot of misunderstandings, and ill-suited governance and regulatory frameworks.
Think of two events that took place last week: 1/ the deal on the EU AI act and 2/ the release of Gemini
5/n
1
0
8
... and most importantly, touching on the nature of intelligence and so-called AGI
2/2
2
0
8
👩👩👧👦 Dans le cadre de son étude sur la détection des mineurs en ligne, le PEReN a développé, en collaboration avec la
@CNIL
, une preuve de concept d'un mécanisme de vérification de l'âge reposant sur une double anonymisation via des vérificateurs tiers👇
0
1
8
Is Gemini a breakthrough ? Likely not.
But how will we know, objectively, when a breakthrough happens ? How can we measure this ?
9/n
1
0
8
What is AGI ?
What are the degrees of intelligence in machines ?
Where are we today?
How fast are we progressing ?
Have we made a leap since ChatGPT?
What are the short and mid-term prospects ?
2/n
2
0
8
@BrndBenjamin
@idrissaberkane
Lire l’article pour éviter de faire des raccourcis à contresens, c’est plus responsable ;)
“Dès que je le pourrai, je le ferai", a-t-il déclaré. Mais il a ajouté qu'il ne voulait pas que les dirigeants "passent devant les autres" et donnent le mauvais exemple.”
5
0
7
Plus, there are a lot of important issues (data contamination, benchmarks robustness…) to solve along the way, while building a solid taxonomy and benchmarks for Machine Intelligence.
12/n
1
0
7
Éclairages intéressants de
@Julien_Pillot
ce matin sur les différents enjeux de réglementation et de gouvernance du métavers à l’occasion de
@VirtualityParis
1
3
8
L'IA est une chance !
L'🇪🇺 et la 🇫🇷 ont l'opportunité de se saisir de, et faire prospérer cet outil, vecteur de connaissance et de créativité.
Très belle conclusion du rapport de la fondation
@robert_schuman
par
@NiccoBianchini
&
@lorenzo_ancona
👏
0
3
8
Implicit acknowledgment that some are pushing against open-source for business reasons, not safety ones👇
🧵Cofounder of
@cohere
@nickfrosst
told me that Cohere protects its LLM weights because of business risks, not the fear of them getting into the wrong hands. “I do want to make the distinction…I’m talking about large language models,” he said. /1
2
0
6
0
1
8
Hate speech regulation: « the quicker and the more ambitious the EU proposals [aka
#DSA
] will be, the more inclined France will be not to take initiatives » said 🇫🇷 Junior Digital Minister
@cedric_o
France leaves door open for new hate speech law before the Digital Services Act, Junior Digital Minister Cédric O told reporters today. With
@vmanancourt
2
8
20
0
4
7
Qui a financé cette campagne ? Mystère !
😶 L'entité ayant payé la campagne n'est pas renseignée
❌ Et surtout cette campagne n'est pas présentée comme une publicité politique (normalement interdite par X au sein de l'UE)...
4/n
1
1
7
@DFintelligence
Il y a cet éclairant fil de
@Dorialexander
via l’angle du copyright.
A short thread on the long and complicated history of AI art and copyright. This is Quadratic Gaussian, a piece of digital art generated by Michael Noll exactly 60 years ago (August 28th 1962) and successfully copyrighted two years later in 1965 after a lengthy debate.
5
32
100
2
0
7
Insightful read on LLM evaluation by
@clefourrier
🤗
clearly explaining existing methodologies and their limitations.
2
2
7
🤝Faire de l'IA générative une chance pour le plus grand nombre, tel est l'objectif des ateliers
@MetaFrance
x
@simplonco
Très heureux d'avoir lancé cette session à Lille aux côtés de
@fbardeau
devant une salle comble !
1
3
7
📊Une étude très éclairante par
@LeCNC
sur les technologies immersives en 🇫🇷
🚶♀️presque 50% des français ont testé la VR
😀 un taux de satisfaction de 93% des expériences VR
🥽 un taux d'équipement de 43% chez les 18-34ans
2
2
6
1/ With the AI act, legislators decided to impose stricter rules to the more capable General-purpose AI model (GPAI).
They agreed to distinguish between « regular » GPAI and GPAI with « systemic risk » on the basis of the compute used for training (10*25 FLOPs threshold)
6/n
1
0
7
Our 2 cents about online hate speech regulation and platform liability in the 🇪🇺 with some 🇩🇪 & 🇫🇷 perspectives, and what this means for the
#DSA
.
With my wonderful colleagues
@DuplanLea
and Melanie Ohnemus !
0
3
7
Camions dans les rues de Paris et Bruxelles, publicités en ligne, achat d'espace média, etc : la campagne menée par l'obscur Control AI qui a ciblé
@cedric_o
et plus largement les fournisseurs de modèles d'IA n'a pas fait l'objet d'un seul article, ni d'une seule enquête...
1
4
6
L’AI Act s’articule autour de la réglementation par le risque; + l’usage est risqué, + il est réglementé.
Mais pour les IA “généralistes” les co-législateurs sont tentés de suivre une approche maximaliste, découplé du degré de risque, ce qui inquiète l’écosystème 🇪🇺👇
The EU’s new AI law will slow down Europe's attempts to adopt generative AI, say founders. “We need to play catch up and now we’ll be doing it with a handicap,” says Jonas Andrulis, CEO Aleph Alpha, a European rival to OpenAI.
@timmpsmith
@siftedeu
4
13
23
1
1
7
Just finished reading a few chapters.
It’s super comprehensive, accessible and well-structured.
Highly recommended for anyone working on AI policy & ethics.
👏thanks for this great work
@IasonGabriel
@Arianna_Manzini
Geoff Keeling & al
1. What are the ethical and societal implications of advanced AI assistants? What might change in a world with more agentic AI?
Our new paper explores these questions:
It’s the result of a one year research collaboration involving 50+ researchers… a🧵
30
199
620
0
0
7
Want to know who is involved in
#AI
governance and ethics shaping ?
Have a look at this excellent mapping of international actors by
@Ph_Lorenz
for
@snv_berlin
#AIethics
#AIpolicy
#AIgovernance
0
2
6
👩🎓Conférence “IA générative et concurrence” par la Chaire Concurrence et Économie Numérique de
@AssasUniversite
1
0
6
Pour
@mathieuweill
la question autour de l’AI Act est de savoir si nous voulons une directive e-commerce de l’IA ou un RGPD de l’IA.
0
3
6
La réalité virtuelle et la réalité augmentée sont aussi réelles et impactantes que notre monde physique.
Telle est la thèse du philosophe
@davidchalmers42
dans ce brillant ouvrage qui offre des réflexions vertigineuses sur les technologies immersives et notre futur.
1
0
6
Open foundation models :
- catalyse innovation
- improve transparency
- combat market concentration
- do not pose greater risks than closed models based on current evidence
These are the takeaways from the latest
@StanfordHAI
report on the governance of open models👇
📢 Are open foundation models actually more risky than closed ones? I'm part of a Stanford-Princeton team that set out to review the evidence and make recommendations for policy makers. Just released:
5
74
241
1
0
6
Bringing political campaigning to the next-level !
A better world is not only possible, it is within our reach.
From mutual aid in our communities to solidarity in the streets, the people of New York have already started to reshape our future.
It’s time to bring the movement to the voting booth.
VOTE this Tuesday, June 23rd.
16K
26K
138K
0
0
6
La
#VR
et l'
#AR
représentent la majeure partie des investissements Reality Labs de Meta explique
@boztank
à
@LUsineDigitale
. Horizon Worlds en est la portion congrue. A bon entendeur !
#M
étavers
1
0
6
Apple overtook Microsoft’s market cap in 2010.
Is Microsoft taking over Apple now, the sign that the last 15 years of shift to mobile is giving way to an AI-first Internet ?
0
2
6
”Pour avoir un rapport efficace au monde, l’être humain doit donc être inscrit dans une communauté de sens, soudée par cette référence commune. Autrement dit le lien social a toujours une structure ternaire, et non simplement binaire. Cette structure est d’abord celle du langage”
Qu'est-ce qu'un monde où l'on ne peut plus se fier à la parole d'autrui ? Plus la valeur de la parole se dégrade, plus nous peinons à « faire société ». Il faut retrouver dans la cité les conditions du parler vrai.— Une pièce de doctrine d'Alain Supiot.
1
22
37
0
1
6
Mettre l'utilisateur au centre de la gouvernance des plateformes numériques; au-delà de l'utopie, voilà une idée stimulante à l'heure des débats sur le web3.0 et la décentralisation. Par
@hisaac25
et Louis-Victor de Franssu
Face au pouvoir des grandes plates-formes numériques, Henri Isaac et Louis-Victor de Franssu, de
@RNumerique
appellent à remettre l’utilisateur au centre de leur régulation afin de mieux l’adapter à l’évolution rapide des usages.
1
4
8
0
1
6
The fight for open source Generative AI
A thoughtful article by competition expert
@ProfSchrepel
spanning many aspects of the issue (AI safety, general regulation, regulatory capture) through the competition lens
0
3
5
🤗Welcome CommonCorpus !
📚500B words of quality and multilingual data, making this the largest public domain corpus to date.
👏Congrats
@ana_stasenko
,
@Dorialexander
and all the people involved !
0
0
6
📡 Qu'ont répondu les différents acteurs en réponse à la consultation de la Commission européenne 🇪🇺 sur l'introduction d'une redevance réseau (
#networkfees
ou
#fairshare
) réclamé par les telcos ? 🔍
Spoiler: beaucoup de monde est contre👇
Un 🧵
1
3
6
The Germany Ministry of Finance
@BMF_Bund
just released a position paper on the digital, programmable euro👇
A big step to foster digital finance and innocation in the EU !
2
3
4