Alan Cowen
@AlanCowen
Followers
2,521
Following
216
Media
93
Statuses
1,338
CEO @hume_ai , teaching AI to make people happy. AI researcher + emotion scientist. Prev @googledeepmind
Joined February 2015
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
#ปิ่นภักดิ์EP8
• 588470 Tweets
TLP HOT WRIST INTERTWINED
• 583230 Tweets
Arsenal
• 222653 Tweets
ارسنال
• 203025 Tweets
#GameofSixTONES
• 156840 Tweets
Lando
• 88942 Tweets
Norris
• 77130 Tweets
Manchester City
• 47730 Tweets
Haaland
• 47340 Tweets
#Number_1
• 46121 Tweets
BBS KAI
• 42224 Tweets
Man City
• 35001 Tweets
SPILL THE FEELS NEW VER 01
• 32063 Tweets
Ricciardo
• 30373 Tweets
McLaren
• 29166 Tweets
Recipe
• 28837 Tweets
ツイステ
• 25976 Tweets
Rodri
• 23602 Tweets
Calafiori
• 22244 Tweets
#MCIARS
• 21253 Tweets
Happy Fall
• 21101 Tweets
Leverkusen
• 16809 Tweets
Fate復刻
• 14145 Tweets
Pinned Tweet
Today we launched EVI 2, a voice-to-voice foundation model
Try it here:
Introducing Empathic Voice Interface 2 (EVI 2), our new voice-to-voice foundation model. EVI 2 merges language and voice into a single model trained specifically for emotional intelligence.
You can try it and start building today.
47
187
977
5
4
39
@paulg
@markessien
this test is mostly pseudoscience. it does, however, tell you whether the explanations you weave for a highly isolated/ambiguous expression are similar to the norm, which you might notice varies substantially by perceived gender
12
3
317
@paulg
I can’t imagine apps that compel young children to post images of themselves publicly and get evaluated every day by thousands of people they barely know would have anything to do with this
5
6
277
@paulg
Until recently social science was mostly a process of making up just-so stories; the uprooting of these kinds of approaches by data-driven science is an exciting shift
8
8
256
Voice interfaces are inevitable. Speech is 4x faster than typing; frees up the eyes & hands; carries more info in its tune, rhythm, & timbre. So we built the first AI with EQ to understand the voice beyond words. It can better predict when to speak, what to say, & how to say it.
Meet Hume’s Empathic Voice Interface (EVI), the first conversational AI with emotional intelligence.
161
491
3K
18
17
195
@ylecun
But the human brain was shaped by many, many trillions of “tokens” over the course of human evolution
4
4
151
@jackclarkSF
This kinda devalues the word "alignment" and feeds into the idea that it's a euphemism for censorship
16
2
105
@tegmark
LLMs said stuff like this years ago (remember Blake Lemoine?). Since then we’ve even been training them on articles about LLMs saying stuff like this
1
0
97
Today, we announced the next big leap on our journey to optimize AI for human well-being. We’re launching a new flagship product, an empathic voice interface (EVI) that can be integrated into any application, that you can try right now. We also raised a $50M Series B
Hume has raised $50 million in funding at a $219 million valuation to build “empathic” conversational AI that can detect over 24 emotions lacing a person's voice.
7
23
67
10
7
99
@NPCollapse
If someone asks a weird question the natural response is to say “is this a test?” That doesn’t require any emergent capability we didn’t already know about
6
0
88
@Nancy_Kanwisher
This is questionable scholarship with a lot of dubious assertions about what people generally mean by the phrase
1
0
68
@dwarkesh_sp
@leopoldasch
I'm not sure if this kind of messaging is good for the world. It's a bit of a self-fulfilling prophecy if people are persuaded by it
6
0
65
@Noahpinion
Kids are being entered into a contest where their images are evaluated by hundreds of acquaintances and strangers every day
0
0
59
Voice AI can show you things and see what you’re seeing while talking to you, freeing the UI (and your eyes) from static text. That means it can enhance the interface of any app, even a website
Hume's EVI API can do more than just chat - it can also take actions! In fact, we embedded EVI as a widget on our website to help you navigate around.
See it in action →
20
33
275
4
2
60
@ylecun
Infants begin life with robust pretrained embeddings for vision (shapes/objects), audition, taste, smell, touch, facial identity/expression, phonetics, syntax, social structures, key affordances (e.g., height, toxic substances, animals), problem-solving strategies, and more
3
2
49
@bkeithpayne
@DEMelnikoff
By default, editors should be trained to see these as a double negative that becomes positive
0
0
49
If you’ve read psychology research that said your common sense intuitions about human emotion were totally wrong, it’s likely that the research was wrong and your intuitions were mostly right, and subsequently borne out by bigger better studies
It seems like the core message of the past decade or so is that this is not the case
If you think about something that seems obviously to be happening, the truth will almost always be: it is totally happening. What we need is just a deeper understanding of *why* it is happening
7
2
64
3
10
49
@paulg
If smaller LLMs trained on less data can perform on par with GPT-4, then clearly the biggest bottleneck to improving LLMs isn’t data or compute—it’s fine-tuning to understand what makes a response “good”
6
5
48
@_akhaliq
A book is <200k tokens. So we’re looking at models that’ll be able to read the first 5 books in a series and write the next one
1
0
48
Have been shocked at how much time people spend talking to our demo: ~10 minutes *on average* (>30 conversational turns). If you wanna build a voice AI people actually enjoy talking to, this is the way to do it
The EVI API is finally here! With our demo alone, we were surprised so many people felt a connection to the world’s first emotionally intelligent voice AI:
✨ ~100K conversations
⏱️ 10 min average conversation length
💬 3M user messages
Start building applications your users
13
80
424
7
1
47
When someone asks you a weird question, your reaction is probably “is this a test”? LLMs are now smart enough to do the same thing, which means it’s time to borrow tactics from psychology to design tests that don’t make it too obvious what they’re testing
Fun story from our internal testing on Claude 3 Opus. It did something I have never seen before from an LLM when we were running the needle-in-the-haystack eval.
For background, this tests a model’s recall ability by inserting a target sentence (the "needle") into a corpus of
577
2K
12K
5
5
43
@NPCollapse
so apparently GPT-3.5 can do this too, just doesn’t do it spontaneously
@mxmmall
@AlanCowen
@NPCollapse
Even GPT-3 can do this. Sure, you have to ask, but isn't "knows its being tested but doesn't volunteer that information unless asked" even more scary if you actually believe this is a dangerous capability?
2
3
32
2
1
42
@ylecun
These studies are confounded. Countries adopting social media later were still developing; we'd expect to see increases in well-being with development (we don't). Individuals who use social media likely have more friends; we'd expect them to be happier as a result (they aren't)
4
1
42
Excited to share that
@hume__ai
just raised a $12.7M Series A led by
@usv
!
It's a privilege to be partnering with investors who share our vision for the future of AI: technology that can align itself with our values based on a scientific understanding of how we express ourselves
Our investment in Hume AI:
"As AI technology begins to shape every aspect of our lives, can we ensure that it cultivates our emotional well-being as a fundamental, overriding objective?"
15
8
89
1
3
41
For those unaware of the main flaw in Yudkowsky's argument: his concerns hinge on "AI foom," the idea that AI beyond a certain threshold of intelligence will figure out how to self-improve, leading to a vicious cycle and the sudden emergence of superintelligence. This overlooks
In the past I have given credit where it was due to Eliezer Yudkowsky for not explicitly advocating violent solutions to the problems with AI development that by his own admission only he and a few other (mostly nontechnical) people see on the horizon
He crossed that line today
185
185
1K
15
3
40
@perrymetzger
So I guess we have time to work on alignment instead of throwing our hands in the air and making futile gestures at stopping everyone on earth from training models
5
3
38
@AISafetyMemes
We don’t have to speculate on this for long; there are open source models just as powerful as Claude and still zero catastrophes
2
0
35
@matspike
This is all missing the point of synthetic data. The point is generally to allow the AI to interact with a system and receive feedback; for instance, generate a million programs, run them, and see which ones worked and which ones didn’t
5
1
32
@rcalo
Why do some of the very people who work on nuclear energy sound the alarm about nuclear safety issues?
11
1
31
@Bob_cart124
But if taken literally it’s a call for conquest, like saying “from Crimea to Lviv, Russia will be free”
5
3
29
@jamiequint
@paulg
@markessien
yep, if you can always predict the normative perception of an expression, that means something. it just doesn't necessarily mean you're good at reasoning about people's emotions in everyday life.
10
0
29
Great explainer on our mission to optimize AI for all signs of human happiness
Is AI's next big leap understanding emotion? $50M for Hume says yes
0
8
16
3
0
30
In this debate with Russell, Bengio, and
@TonyZador
,
@ylecun
makes a key point. They all basically agree on the missing ingredient in AI safety: updatable objective functions, with the ability to correct harmful effects in production (presumably using harm/well-being measures)
@rich_gast
Scientific debates are good.
Here is an example below.
But debating AI ethics and safety with extreme AI doomers is like debating evolution with creationists.
Pointless.
15
26
164
4
6
29
Social media algorithms have long been misaligned: engagement≠user satisfaction.
An important paper by
@william__brady
@josh_c_jackson
et al drives this point home.
We should be optimizing directly for user satisfaction.
0
3
26
@cremieuxrecueil
@paulg
Smart people get less worked up about dumb things, like psychology stimuli. They probably get more emotional about complex ideas. “Reason is, and ought only to be the slave of the passions, and can never pretend to any other office than to serve and obey them.”
1
1
27
@JMannhart
If the timeline is 20 years, or even 5, what does a 6 month moratorium on GPT-5 buy us? Does it delay the problem by 6 months, or not at all? Does it get us 6 months closer to a solution, or 6 months further?
8
0
26
@WarClandestine
Speaking as someone who actually knows how these models work--no, it doesn't mean Gemini was taught this. It means Gemini was fine-tuned to speak in a certain style, probably the style of someone excessively woke, and it overgeneralized.
3
1
25
Watch EVI roast
@Jason
New Episode!
@alancowen
breaks down
@hume_ai
's new Empathic Voice Interface (EVI) and the AI's ability to understand and respond to human emotions
-- Demo of Hume's EVI
-- Demo of Hume's Measurement API
-- AI comedy roast of JCal
-- The future implications of this technology
239
7
33
2
0
25
@RitchieTorres
If your goal is peace with your neighbor, why use a motto that your neighbor interprets as a threat?
15
1
24
@DanHendrycks
@jeremyphoward
It’s open source licensed; the main reason for the patent may be defensive (to establish their right to permissively license the technology)
2
0
25
@ineshipolito
They should’ve gone further: all scientific theories of consciousness are pseudoscience
2
1
24
@ylecun
There are a lot more bots on this platform passing as human and building credibility, and it’s unclear how much of a threat they’ll pose, but it took a couple more generations of improvement to get here
3
0
24
@martinmbauer
Science also operates a lot more efficiently than people realize (very smart people being paid very little)
0
2
23
@emollick
Ironically, the AI race has actually slowed down the development of frontier models, because the biggest players are no longer publishing their best ideas
0
1
23
Pretty cool to see
@hume_ai
in the
@CBinsights
#AI100
, but "emotion analytics" is just the first step. At Hume, we’ve been pretraining and aligning LLMs with human expressions so they can learn people’s implicit preferences and self-improve over time.
Think about everything
0
6
22
@MerGarza
@Psicologia_UANL
Y mi emoción favorita que falta en la película, el asombro (imaginado por la IA)
1
1
22
ChatGPT recognizes its own biggest weakness: the need for empathy
4
7
20
Enjoyed participating in the Ethical Affective Computing panel
@acii_conf
today
Favorite question from Q&A: “where’s the line between personalization and manipulation?”
A: personalization means optimizing for your preferences; manipulation means optimizing your behavior for
Academic and industry experts share thought-provoking considerations and research directions for building ethical affective computing systems, today at the panel session in
#ACII2023
.
0
1
31
1
2
20
@paulg
It was relatively recent that Google merged with Deepmind, so we don’t know how that will play out, and in the meantime some of the expert models they’ve been building (Med-PaLM 2, Robotic Transformer 2) are state of the art
4
0
21
@juang_acosta
@joecappadona
@ylecun
The human brain is not a blank slate, but a heavily pretrained architecture
1
0
21
@AISafetyMemes
@ylecun
It seems irresponsible for an expert to propagate the dangerous misconception that this thing may be conscious or self-aware when the evidence is so obviously lacking (it's repeating the same tropes that Blake Lemoine squeezed out of Lambda 2 years ago)
1
0
21
@JustChattingEh
@Mikel_Jollett
@terrifyingwoman
No they’ve diverted materials meant for building civilian infrastructure. They have lots of buildings…
1
0
19
@varun_mathur
A little dramatic; they could probably still raise at a $40B valuation without much effort
3
0
19
1
0
18
Had a great conversation with
@mignano
on
@lightspeedvp
’s podcast, Generative Now. We spoke about the importance of emotional intelligence in AI, the science behind Hume's Empathic Voice Interface, and why AGI would be useless if we built it without an added layer of EQ.
New episode of the pod: I spoke with
@AlanCowen
, CEO of
@hume_ai
, the company building the world's first empathic AI.
Unlike other conversational AI's, Hume can actually understand users' emotions just by analyzing facial expressions and tone of voice, which helps them deliver a
2
2
18
0
2
18
The future of AI is bringing your personal AI to an application, recognizing it by voice, trusting it to fulfill your preferences as it helps you navigate, with much of the UI being generated on the fly.
Part of a fascinating conversation about the future of voice and music AI
Our CEO
@AlanCowen
on why emotionally intelligent voice interfaces are the key to AI that understands our needs.
1
8
50
1
0
17
@gfodor
Yep, this is just the LLM version of the invisible gorilla test (and all those other tasks that cause the human brain to “glitch”)
0
0
17
Had a great conversation with
@danshipper
. We got pretty deep into the science behind Hume, including some spicier takes on the history of emotion science and where the field will go next. (Minor clarification: some people take the phrase “read emotions” literally. You can’t read
This AI can read emotions better than you can.
It was created by Hume (
@hume_ai
) an AI research lab developing models that can read your face and your voice with uncanny accuracy. Their hope is that models that can read your emotions will help create AI that optimizes for human
20
32
240
1
1
17
If we build AI right, we’ll never run out of jobs for humans. We’ll have full employment where everyone provides human feedback for AI through an incredibly fulfilling dialectical process, and the AI uses this feedback to continuously improve its ability to make humans happy
3
1
14
@hamandcheese
@sama
If you replace “the thing he’s trying to create” with “the thing people will inevitably create” it really changes the tone of the question
1
0
17
@abacaj
This seems a bit misleading. The results show that transformers can generalize beyond their training data as long as the data is drawn from a function class present in the training data, which for LLMs might just mean data that adheres to a unifying world model
0
2
16
Before you can trust AI to make decisions for you, you have to believe it's working on your behalf and not for an organization. But few people can fine-tune LLMs themselves. Helping people align AI to their preferences (without hurting anyone) is the biggest opportunity around
At least I can align a local LLM to me and not some organization or whatever goals they have set forth. This gives me hope and makes sense in the long term, it’s not always obvious how the big hosted models are aligned, to who or what beliefs
7
6
95
0
2
17
This is genuinely fun to play with
What happens when EVI is customized to act like… a wistful fridge?🤔or… a jealous houseplant? or literally anything else you can imagine? Our Empathic Voice Interface playground is now live!
35
40
400
2
0
17
EVI is a new AI system that understands and generates expressive speech, trained on millions of human conversations. We’re still ironing out some kinks, but it’s already pretty fun to talk to—you can “feel the EQ.”
It can be integrated into any application with just a few lines
1
1
14
seems unintuitive at first, especially for coding, but speech will be a more important AI interface than text. it’s faster and AI doesn’t need polished/copy-edited instructions so the advantages of writing fade away
I've been really enjoying editing code by voice: using
@cursor_ai
for IDE +
@superwhisperapp
for transcription.
It's so fast! Hit cmd-k, say what I want, boom. Faster than typing it myself. The prompts can get quite detailed too, voice makes it fast enough
23
88
758
2
0
16
@hume__ai
published a paper in Nature Human Behavior! It introduces the first AI model that captures the complexity and nuance of human vocal expression across cultures.
Why this is important for the future of AI and how it learns to serve human needs:
2
1
16
It’s been great for
@hume_ai
being part of and working with
@natfriedman
,
@danielgross
, and
@Hersh_Desai
Applications are open for batch 3 of for pre-seed and seed-stage companies building AI products! Deadline is Feb 16.
As an experiment, this batch we are offering the option of either receiving $250k on an uncapped note, or $2.5M at a $25M cap.
58
155
1K
0
2
15
@emollick
It’s impossible for these lines to be linear since MMLU maxes out at 100 (or ~95 depending on who you ask), so their convergence may be illusory
0
0
15
@sama
A lot of people are confused about the word "align." We should use simpler language, like "optimize in favor of doing good and against doing harm" (where "doing good" means intent completion, preference satisfaction, happiness, interest, well-being, mental health, etc.)
2
1
15
@sama
As an academic: academy isn’t collapsing. It’s just always been semi-dysfunctional. But if anything, it’s getting better (faster, more rigorous, more open, more diverse, more integrated with industry)
1
0
15
@ID_AA_Carmack
But the human brain was shaped by many, many trillions of words over the course of human evolution
2
0
15
The
@hume_ai
team is really looking forward to working with
@natfriedman
,
@danielgross
, and
@Hersh_Desai
!
1
3
15
Alignment is feasible because the parameter space of language (and human behavior more broadly) is dense with affordances for human emotion and well-being—“the mapping from parameter space to goal space is extremely compressive”
@liron
@mezaoptimizer
I'd be fine with doing a podcast. I think the crux of our disagreement is pretty clear, though. You seem to think there are 'basic principles of “optimization theory”' that let you confidently conclude that alignment is very difficult. I think such laws, insofar as we know enough
9
4
52
1
2
13
is a bold but long-awaited endeavor for my pursuit of AI over 4 decades. Proud to introduce the world's top open-source model Yi-34B as our first release to the developers' community encouraging fantastic LLM projects with a moderate-size, high-performing
65
187
1K
0
3
15
It's great that the highest-grossing film of the year is a lesson in emotion science and Humean philosophy ("Reason is and ought only to be the slave of the passions")
Up there with the best Pixar films imo
‘Inside Out 2’ becomes the highest-grossing film of 2024 at the domestic box office with $285.7 million.
123
544
9K
3
0
15
@Carnage4Life
“An AI that cares about understanding the universe is unlikely to annihilate humans because we are an interesting part of the universe.” Why not just directly optimize the AI to care about humans?
2
4
15