Gerard de Melo
@gdemelo.bsky.social
Professor for AI at Hasso Plattner Institute and University of Potsdam
Berlin (prev. Rutgers NJ USA, Tsinghua Beijing, Berkeley)
http://gerard.demelo.org
Berlin (prev. Rutgers NJ USA, Tsinghua Beijing, Berkeley)
http://gerard.demelo.org
“When it comes to AI, European companies rarely play in the top league” – My perspective on why we need AI sovereignty. 👉🏽 hpi.de/en/article/w...
🚨Join us at the AI @ HPI Conference 2025 this December:
hpi.de/en/ai-hpi-co...
🚨Join us at the AI @ HPI Conference 2025 this December:
hpi.de/en/ai-hpi-co...
October 30, 2025 at 12:31 PM
“When it comes to AI, European companies rarely play in the top league” – My perspective on why we need AI sovereignty. 👉🏽 hpi.de/en/article/w...
🚨Join us at the AI @ HPI Conference 2025 this December:
hpi.de/en/ai-hpi-co...
🚨Join us at the AI @ HPI Conference 2025 this December:
hpi.de/en/ai-hpi-co...
Remembering Pushpak Bhattacharyya, known as Godfather of Indian NLP, former President of ACL + Director of IIT Patna.
Despite his countless achievements, he remained humble and kind. He was one of the first people I met at an academic conference.
May he rest in peace. www.cse.iitb.ac.in/memorial
Despite his countless achievements, he remained humble and kind. He was one of the first people I met at an academic conference.
May he rest in peace. www.cse.iitb.ac.in/memorial
October 6, 2025 at 10:58 AM
Remembering Pushpak Bhattacharyya, known as Godfather of Indian NLP, former President of ACL + Director of IIT Patna.
Despite his countless achievements, he remained humble and kind. He was one of the first people I met at an academic conference.
May he rest in peace. www.cse.iitb.ac.in/memorial
Despite his countless achievements, he remained humble and kind. He was one of the first people I met at an academic conference.
May he rest in peace. www.cse.iitb.ac.in/memorial
OpenAI's Benedict Kerres at our @hpi.bsky.social Tech Leadership Conference:
ChatGPT used for relationship and health advice a lot in Germany
ChatGPT used for relationship and health advice a lot in Germany
July 18, 2025 at 10:24 AM
OpenAI's Benedict Kerres at our @hpi.bsky.social Tech Leadership Conference:
ChatGPT used for relationship and health advice a lot in Germany
ChatGPT used for relationship and health advice a lot in Germany
Papers *written* and *reviewed* by AI 🤖
The Stanford Agents4Science conference is seeking research papers mainly authored by AI, which will then be peer-reviewed by multiple AI systems.
Might seem odd to us now, but this might become common practice in the future.
agents4science.stanford.edu
The Stanford Agents4Science conference is seeking research papers mainly authored by AI, which will then be peer-reviewed by multiple AI systems.
Might seem odd to us now, but this might become common practice in the future.
agents4science.stanford.edu
July 13, 2025 at 12:47 PM
Papers *written* and *reviewed* by AI 🤖
The Stanford Agents4Science conference is seeking research papers mainly authored by AI, which will then be peer-reviewed by multiple AI systems.
Might seem odd to us now, but this might become common practice in the future.
agents4science.stanford.edu
The Stanford Agents4Science conference is seeking research papers mainly authored by AI, which will then be peer-reviewed by multiple AI systems.
Might seem odd to us now, but this might become common practice in the future.
agents4science.stanford.edu
😮 Recent OpenAI / former DeepMind researcher talking about "mass migration as being worse for Europe than WW2 was" and extremist phantasies such as "large-scale remigration".
x.com/RichardMCNgo...
x.com/RichardMCNgo...
June 22, 2025 at 1:09 PM
😮 Recent OpenAI / former DeepMind researcher talking about "mass migration as being worse for Europe than WW2 was" and extremist phantasies such as "large-scale remigration".
x.com/RichardMCNgo...
x.com/RichardMCNgo...
Huh, but that's a picture of the cookie monster? My slides for Tuesday happen to include ELMo (red) and BERT (yellow) :)
May 4, 2025 at 10:27 AM
Huh, but that's a picture of the cookie monster? My slides for Tuesday happen to include ELMo (red) and BERT (yellow) :)
We can train LLMs to say "I don't know" when they make mistakes. See: proceedings.neurips.cc/paper_files/...
April 23, 2025 at 5:18 PM
We can train LLMs to say "I don't know" when they make mistakes. See: proceedings.neurips.cc/paper_files/...
Even simpler: Encourage everyone to end their talk with a beautiful Bluesky QR code
qrsurge.com/bluesky-qr-c...
qrsurge.com/bluesky-qr-c...
March 27, 2025 at 2:44 PM
Even simpler: Encourage everyone to end their talk with a beautiful Bluesky QR code
qrsurge.com/bluesky-qr-c...
qrsurge.com/bluesky-qr-c...
March 27, 2025 at 9:08 AM
🤯
An AI superintelligence strategy
involving deterrence based on sabotage, espionage, etc.
plus nonproliferation and competitiveness via military + other strength
proposed by
ex-Google CEO Eric Schmidt
with Dan Hendrycks (Berkeley)
and Alexandr Wang (Scale AI)
at www.nationalsecurity.ai
An AI superintelligence strategy
involving deterrence based on sabotage, espionage, etc.
plus nonproliferation and competitiveness via military + other strength
proposed by
ex-Google CEO Eric Schmidt
with Dan Hendrycks (Berkeley)
and Alexandr Wang (Scale AI)
at www.nationalsecurity.ai
March 6, 2025 at 11:53 AM
🤯
An AI superintelligence strategy
involving deterrence based on sabotage, espionage, etc.
plus nonproliferation and competitiveness via military + other strength
proposed by
ex-Google CEO Eric Schmidt
with Dan Hendrycks (Berkeley)
and Alexandr Wang (Scale AI)
at www.nationalsecurity.ai
An AI superintelligence strategy
involving deterrence based on sabotage, espionage, etc.
plus nonproliferation and competitiveness via military + other strength
proposed by
ex-Google CEO Eric Schmidt
with Dan Hendrycks (Berkeley)
and Alexandr Wang (Scale AI)
at www.nationalsecurity.ai
Warfare: AI sadly already a major aspect
• Millions of drones used by Ukraine
• AI for target recommendations
• AI piloting is the future
• Air defense cannot wait for a human in the loop
US General David Petraeus at #AIActionSummit w/ Clara Kaluderovic who develops mental health AI for Ukraine
• Millions of drones used by Ukraine
• AI for target recommendations
• AI piloting is the future
• Air defense cannot wait for a human in the loop
US General David Petraeus at #AIActionSummit w/ Clara Kaluderovic who develops mental health AI for Ukraine
February 13, 2025 at 2:59 PM
Warfare: AI sadly already a major aspect
• Millions of drones used by Ukraine
• AI for target recommendations
• AI piloting is the future
• Air defense cannot wait for a human in the loop
US General David Petraeus at #AIActionSummit w/ Clara Kaluderovic who develops mental health AI for Ukraine
• Millions of drones used by Ukraine
• AI for target recommendations
• AI piloting is the future
• Air defense cannot wait for a human in the loop
US General David Petraeus at #AIActionSummit w/ Clara Kaluderovic who develops mental health AI for Ukraine
Programmer jobs as canary in the coal mine for job displacement?
Traditionally safe career, but special properties – AI solutions testable/verifiable!
Blaise Agüera y Arcas (Google), @erikbryn.bsky.social (@stanfordhai.bsky.social), Terah Lyons (ex PAI), Alexander Vollert (AXA)
#AIActionSummit
Traditionally safe career, but special properties – AI solutions testable/verifiable!
Blaise Agüera y Arcas (Google), @erikbryn.bsky.social (@stanfordhai.bsky.social), Terah Lyons (ex PAI), Alexander Vollert (AXA)
#AIActionSummit
February 12, 2025 at 4:24 PM
Programmer jobs as canary in the coal mine for job displacement?
Traditionally safe career, but special properties – AI solutions testable/verifiable!
Blaise Agüera y Arcas (Google), @erikbryn.bsky.social (@stanfordhai.bsky.social), Terah Lyons (ex PAI), Alexander Vollert (AXA)
#AIActionSummit
Traditionally safe career, but special properties – AI solutions testable/verifiable!
Blaise Agüera y Arcas (Google), @erikbryn.bsky.social (@stanfordhai.bsky.social), Terah Lyons (ex PAI), Alexander Vollert (AXA)
#AIActionSummit
AI optimism vs. pessimism: Despite concerning developments, need to be optimistic in our actions, e.g. need massive compute for academia/civil society
Fei-Fei Li @drfeifei.bsky.social, Rob Reich, Tino Cuéllar (Carnegie), Josephine Teo (SG minister)
AXA @stanfordhai.bsky.social at #AIActionSummit
Fei-Fei Li @drfeifei.bsky.social, Rob Reich, Tino Cuéllar (Carnegie), Josephine Teo (SG minister)
AXA @stanfordhai.bsky.social at #AIActionSummit
February 12, 2025 at 3:02 PM
AI optimism vs. pessimism: Despite concerning developments, need to be optimistic in our actions, e.g. need massive compute for academia/civil society
Fei-Fei Li @drfeifei.bsky.social, Rob Reich, Tino Cuéllar (Carnegie), Josephine Teo (SG minister)
AXA @stanfordhai.bsky.social at #AIActionSummit
Fei-Fei Li @drfeifei.bsky.social, Rob Reich, Tino Cuéllar (Carnegie), Josephine Teo (SG minister)
AXA @stanfordhai.bsky.social at #AIActionSummit
Regulating AI only if technically feasible?
What if someone wants to build a nuclear plant in the middle of Paris and argues it's not technically feasible to ensure its safety?
Stuart Russell, @yoshuabengio.bsky.social, Dawn Song, Irakli Beridze, Xue Lan, Karine Perset
#aisconnect #AIActionSummit
What if someone wants to build a nuclear plant in the middle of Paris and argues it's not technically feasible to ensure its safety?
Stuart Russell, @yoshuabengio.bsky.social, Dawn Song, Irakli Beridze, Xue Lan, Karine Perset
#aisconnect #AIActionSummit
February 11, 2025 at 2:26 PM
Regulating AI only if technically feasible?
What if someone wants to build a nuclear plant in the middle of Paris and argues it's not technically feasible to ensure its safety?
Stuart Russell, @yoshuabengio.bsky.social, Dawn Song, Irakli Beridze, Xue Lan, Karine Perset
#aisconnect #AIActionSummit
What if someone wants to build a nuclear plant in the middle of Paris and argues it's not technically feasible to ensure its safety?
Stuart Russell, @yoshuabengio.bsky.social, Dawn Song, Irakli Beridze, Xue Lan, Karine Perset
#aisconnect #AIActionSummit
Although Bengio @yoshuabengio.bsky.social was here as well, this panel with Danielle Allen (Harvard) and Philippe Aghion (Collège de France/INSEAD) was a lot more optimistic that we can find ways for society to manage the transformative power of AI, as suggested by @erikbryn.bsky.social
February 11, 2025 at 11:23 AM
Although Bengio @yoshuabengio.bsky.social was here as well, this panel with Danielle Allen (Harvard) and Philippe Aghion (Collège de France/INSEAD) was a lot more optimistic that we can find ways for society to manage the transformative power of AI, as suggested by @erikbryn.bsky.social
Erik Brynjolfsson @erikbryn.bsky.social @stanforddel.bsky.social @stanfordhai.bsky.social
We have a long runway:
For many tasks, even when AI outperforms humans, for a while humans in conjunction with machines will still be better.
Need research on how to manage this transition.
#AIActionSummit
We have a long runway:
For many tasks, even when AI outperforms humans, for a while humans in conjunction with machines will still be better.
Need research on how to manage this transition.
#AIActionSummit
February 11, 2025 at 11:15 AM
Erik Brynjolfsson @erikbryn.bsky.social @stanforddel.bsky.social @stanfordhai.bsky.social
We have a long runway:
For many tasks, even when AI outperforms humans, for a while humans in conjunction with machines will still be better.
Need research on how to manage this transition.
#AIActionSummit
We have a long runway:
For many tasks, even when AI outperforms humans, for a while humans in conjunction with machines will still be better.
Need research on how to manage this transition.
#AIActionSummit
Kate Crawford @katecrawford.bsky.social expanding on Meg Mitchell @mmitchell.bsky.social, suggesting further alternative names for #IASEAI25 😁
How about
IASESAPAI - International Association for Safe & Ethical & Sustainable As Possible AI?
#AIActionSummit #IASEAI25 #IASEAI2025
How about
IASESAPAI - International Association for Safe & Ethical & Sustainable As Possible AI?
#AIActionSummit #IASEAI25 #IASEAI2025
February 11, 2025 at 11:03 AM
Kate Crawford @katecrawford.bsky.social expanding on Meg Mitchell @mmitchell.bsky.social, suggesting further alternative names for #IASEAI25 😁
How about
IASESAPAI - International Association for Safe & Ethical & Sustainable As Possible AI?
#AIActionSummit #IASEAI25 #IASEAI2025
How about
IASESAPAI - International Association for Safe & Ethical & Sustainable As Possible AI?
#AIActionSummit #IASEAI25 #IASEAI2025
Thomas Dietterich @tdietterich.bsky.social at #IASEAI25
We need distribution-independent learning and evaluation!
Machine learning models (e.g. in self-driving cars) need to recognize new, unanticipated scenarios.
#IASEAI2025 #AIActionSummit
We need distribution-independent learning and evaluation!
Machine learning models (e.g. in self-driving cars) need to recognize new, unanticipated scenarios.
#IASEAI2025 #AIActionSummit
February 9, 2025 at 1:46 PM
Thomas Dietterich @tdietterich.bsky.social at #IASEAI25
We need distribution-independent learning and evaluation!
Machine learning models (e.g. in self-driving cars) need to recognize new, unanticipated scenarios.
#IASEAI2025 #AIActionSummit
We need distribution-independent learning and evaluation!
Machine learning models (e.g. in self-driving cars) need to recognize new, unanticipated scenarios.
#IASEAI2025 #AIActionSummit
AI doesn't inherently possess a desire to avoid being shut off
@yoshuabengio.bsky.social on how AI self-preservation
could still arise
– humans intentionally add it
– human imitation
– unintentional subgoal
– reward tampering
– competition betw. AI developers
#AIActionSummit #IASEAI25 #IASEAI2025
@yoshuabengio.bsky.social on how AI self-preservation
could still arise
– humans intentionally add it
– human imitation
– unintentional subgoal
– reward tampering
– competition betw. AI developers
#AIActionSummit #IASEAI25 #IASEAI2025
February 9, 2025 at 1:35 PM
AI doesn't inherently possess a desire to avoid being shut off
@yoshuabengio.bsky.social on how AI self-preservation
could still arise
– humans intentionally add it
– human imitation
– unintentional subgoal
– reward tampering
– competition betw. AI developers
#AIActionSummit #IASEAI25 #IASEAI2025
@yoshuabengio.bsky.social on how AI self-preservation
could still arise
– humans intentionally add it
– human imitation
– unintentional subgoal
– reward tampering
– competition betw. AI developers
#AIActionSummit #IASEAI25 #IASEAI2025
AI ethics & AI safety should 𝐧𝐨𝐭 be two separate disciplines/communities!
AI ethics enhances ̄AI robustness, transparency, and accounting for pluralism
🤝
AI safety enhances AI interpretability and control
Anca Dragan (Google DeepMind/Berkeley) at #IASEAI25 #IASEAI2025 #AIActionSummit in Paris
AI ethics enhances ̄AI robustness, transparency, and accounting for pluralism
🤝
AI safety enhances AI interpretability and control
Anca Dragan (Google DeepMind/Berkeley) at #IASEAI25 #IASEAI2025 #AIActionSummit in Paris
February 9, 2025 at 1:27 PM
AI ethics & AI safety should 𝐧𝐨𝐭 be two separate disciplines/communities!
AI ethics enhances ̄AI robustness, transparency, and accounting for pluralism
🤝
AI safety enhances AI interpretability and control
Anca Dragan (Google DeepMind/Berkeley) at #IASEAI25 #IASEAI2025 #AIActionSummit in Paris
AI ethics enhances ̄AI robustness, transparency, and accounting for pluralism
🤝
AI safety enhances AI interpretability and control
Anca Dragan (Google DeepMind/Berkeley) at #IASEAI25 #IASEAI2025 #AIActionSummit in Paris
Meg Mitchell @mmitchell.bsky.social suggests the new name
IAAIHVPVSP for #IASEAI25 😁
International Association for AI Informed by Human Values with Prioritization of the Value of Safety in Particular
#IASEAI2025 #AIActionSummit
IAAIHVPVSP for #IASEAI25 😁
International Association for AI Informed by Human Values with Prioritization of the Value of Safety in Particular
#IASEAI2025 #AIActionSummit
February 6, 2025 at 4:03 PM
Meg Mitchell @mmitchell.bsky.social suggests the new name
IAAIHVPVSP for #IASEAI25 😁
International Association for AI Informed by Human Values with Prioritization of the Value of Safety in Particular
#IASEAI2025 #AIActionSummit
IAAIHVPVSP for #IASEAI25 😁
International Association for AI Informed by Human Values with Prioritization of the Value of Safety in Particular
#IASEAI2025 #AIActionSummit
~1,000 requests to speak at the inaugural Conference of the International Association for Safe and Ethical AI (IASEAI ‘25)!
Not the most memorable name, but it seems we'll need to remember it.
Opening remarks by Stuart Russell, co-author of main AI textbook
#AIActionSummit #IASEAI25 #IASEAI2025
Not the most memorable name, but it seems we'll need to remember it.
Opening remarks by Stuart Russell, co-author of main AI textbook
#AIActionSummit #IASEAI25 #IASEAI2025
February 6, 2025 at 4:01 PM
~1,000 requests to speak at the inaugural Conference of the International Association for Safe and Ethical AI (IASEAI ‘25)!
Not the most memorable name, but it seems we'll need to remember it.
Opening remarks by Stuart Russell, co-author of main AI textbook
#AIActionSummit #IASEAI25 #IASEAI2025
Not the most memorable name, but it seems we'll need to remember it.
Opening remarks by Stuart Russell, co-author of main AI textbook
#AIActionSummit #IASEAI25 #IASEAI2025
How do we safely evaluate *agentic* AI systems?
Rumman Chowdhury @ruchowdh.bsky.social at #AIActionSummit side event argues that:
* real-time log evaluation may fail to prevent serious damage
* sandboxes/synthetic environments are a potential solution, but costly/hard to create. More work needed!
Rumman Chowdhury @ruchowdh.bsky.social at #AIActionSummit side event argues that:
* real-time log evaluation may fail to prevent serious damage
* sandboxes/synthetic environments are a potential solution, but costly/hard to create. More work needed!
February 6, 2025 at 2:13 PM
How do we safely evaluate *agentic* AI systems?
Rumman Chowdhury @ruchowdh.bsky.social at #AIActionSummit side event argues that:
* real-time log evaluation may fail to prevent serious damage
* sandboxes/synthetic environments are a potential solution, but costly/hard to create. More work needed!
Rumman Chowdhury @ruchowdh.bsky.social at #AIActionSummit side event argues that:
* real-time log evaluation may fail to prevent serious damage
* sandboxes/synthetic environments are a potential solution, but costly/hard to create. More work needed!
AI/ML folks still on the other platform
February 5, 2025 at 7:24 PM
AI/ML folks still on the other platform
Lots of random encounters at #Davos2025
For example, walking along the street, you might bump into Steven Pinker @sapinker.bsky.social
For example, walking along the street, you might bump into Steven Pinker @sapinker.bsky.social
January 26, 2025 at 5:42 PM
Lots of random encounters at #Davos2025
For example, walking along the street, you might bump into Steven Pinker @sapinker.bsky.social
For example, walking along the street, you might bump into Steven Pinker @sapinker.bsky.social