Surprising things to know about LLMs - Best Ones:
1. LLMs predictably get more capable with increasing investment, even without targeted innovation.
3. LLMs often appear to learn and use representations of the outside world.
6. Human performance on a task isn’t an upper bound on LLM performance.
Paper
1. LLMs predictably get more capable with increasing investment, even without targeted innovation.
3. LLMs often appear to learn and use representations of the outside world.
6. Human performance on a task isn’t an upper bound on LLM performance.
Paper
👀10👏5❤2👍1💯1
Google finally releases paper on the TPUv4 AI training hardware they’ve been using since 2020
TPU v4 is the fifth Google domain specific architecture (DSA) and its third supercomputer for such ML models. Optical circuit switches (OCSes) dynamically reconfigure its interconnect topology to improve scale, availability, utilization, modularity, deployment, security, power, and performance; users can pick a twisted 3D torus topology if desired. Much cheaper, lower power, and faster than Infiniband, OCSes and underlying optical components are <5% of system cost and <3% of system power. Each TPU v4 includes SparseCores, dataflow processors that accelerate models that rely on embeddings by 5x-7x yet use only 5% of die area and power. Deployed since 2020, TPU v4 outperforms TPU v3 by 2.1x and improves performance/Watt by 2.7x. The TPU v4 supercomputer is 4x larger at 4096 chips and thus ~10x faster overall, which along with OCS flexibility helps large language models. For similar sized systems, it is ~4.3x-4.5x faster than the Graphcore IPU Bow and is 1.2x-1.7x faster and uses 1.3x-1.9x less power than the Nvidia A100. TPU v4s inside the energy-optimized warehouse scale computers of Google Cloud use ~3x less energy and produce ~20x less CO2e than contemporary DSAs in a typical on-premise data center.
Paper
TPU v4 is the fifth Google domain specific architecture (DSA) and its third supercomputer for such ML models. Optical circuit switches (OCSes) dynamically reconfigure its interconnect topology to improve scale, availability, utilization, modularity, deployment, security, power, and performance; users can pick a twisted 3D torus topology if desired. Much cheaper, lower power, and faster than Infiniband, OCSes and underlying optical components are <5% of system cost and <3% of system power. Each TPU v4 includes SparseCores, dataflow processors that accelerate models that rely on embeddings by 5x-7x yet use only 5% of die area and power. Deployed since 2020, TPU v4 outperforms TPU v3 by 2.1x and improves performance/Watt by 2.7x. The TPU v4 supercomputer is 4x larger at 4096 chips and thus ~10x faster overall, which along with OCS flexibility helps large language models. For similar sized systems, it is ~4.3x-4.5x faster than the Graphcore IPU Bow and is 1.2x-1.7x faster and uses 1.3x-1.9x less power than the Nvidia A100. TPU v4s inside the energy-optimized warehouse scale computers of Google Cloud use ~3x less energy and produce ~20x less CO2e than contemporary DSAs in a typical on-premise data center.
Paper
👍5❤1👌1
This media is not supported in your browser
VIEW IN TELEGRAM
Soon: Runway Text-to-Video AI
“These visuals were generated using only text prompts. Early access rolling out this week.”
Paper
Website
“These visuals were generated using only text prompts. Early access rolling out this week.”
Paper
Website
👍17🫡5❤1
Using text chat to edit videos:
Structure and Content-Guided Video Synthesis with Diffusion Models
Text-guided generative diffusion models unlock powerful image creation and editing tools. While these have been extended to video generation, current approaches that edit the content of existing footage while retaining structure require expensive re-training for every input or rely on error-prone propagation of image edits across frames.
In this work, we present a structure and content-guided video diffusion model that edits videos based on visual or textual descriptions of the desired output.
Paper
Structure and Content-Guided Video Synthesis with Diffusion Models
Text-guided generative diffusion models unlock powerful image creation and editing tools. While these have been extended to video generation, current approaches that edit the content of existing footage while retaining structure require expensive re-training for every input or rely on error-prone propagation of image edits across frames.
In this work, we present a structure and content-guided video diffusion model that edits videos based on visual or textual descriptions of the desired output.
Paper
👍8❤2
User 1: ChatGPT produces code that has bugs
User 2: Did you try asking ChatGPT to find the bug?
User 1: Well I’ll be damned
Counterintuitively, for LLMs — splitting into multiple interactive generate+check+fix chat steps, rather than of trying to do it all in 1 step, is distinctly more powerful.
Eventually we’ll have solid explainations as to why this is. Until then, remember, 2+ chat steps is often far better than 1.
User 2: Did you try asking ChatGPT to find the bug?
User 1: Well I’ll be damned
Counterintuitively, for LLMs — splitting into multiple interactive generate+check+fix chat steps, rather than of trying to do it all in 1 step, is distinctly more powerful.
Eventually we’ll have solid explainations as to why this is. Until then, remember, 2+ chat steps is often far better than 1.
👍16👌3👏2❤1🙏1
The Anthropomorphization Fallacy Fallacy
= When you're so stupid that you believe simply by throwing around the word “anthropomorphization”, said in an insulting tone, that you've made a real argument.
These anthropomorphization-ers never do seem to explain why all human-AI analogies must be so terrible. Purely a feels argument.
So poor an argument, it’s not even wrong.
As those who’ve closely followed deep learning since the pre-2011 era could tell you, the anthropomorphization fallacy is an absolute lie.
Deep learning is shockingly similar to humans in its emergent behavior, with more parallels by the day.
Not matching in low-level implementation, but so much of the high-level behavior rhymes.
Totally different ways that humans and AIs implement their neural functions, and you won’t figure out much about how to create the next AI by studying cells under a microscope, but the emergent behavior parallels arising from each keeps growing by the day.
(In fact, the direction of discovery has reversed, with AI discoveries now often happening first, with later confirmation of biological systems doing the same happening much later.)
Attention, numerous reinforcement learning reward system concepts, the full list is huge.
AI Anthropomorphization fallacy fallacy. I.e. “dumb AAFF”.
= When you're so stupid that you believe simply by throwing around the word “anthropomorphization”, said in an insulting tone, that you've made a real argument.
These anthropomorphization-ers never do seem to explain why all human-AI analogies must be so terrible. Purely a feels argument.
So poor an argument, it’s not even wrong.
As those who’ve closely followed deep learning since the pre-2011 era could tell you, the anthropomorphization fallacy is an absolute lie.
Deep learning is shockingly similar to humans in its emergent behavior, with more parallels by the day.
Not matching in low-level implementation, but so much of the high-level behavior rhymes.
Totally different ways that humans and AIs implement their neural functions, and you won’t figure out much about how to create the next AI by studying cells under a microscope, but the emergent behavior parallels arising from each keeps growing by the day.
(In fact, the direction of discovery has reversed, with AI discoveries now often happening first, with later confirmation of biological systems doing the same happening much later.)
Attention, numerous reinforcement learning reward system concepts, the full list is huge.
AI Anthropomorphization fallacy fallacy. I.e. “dumb AAFF”.
👍6🤯3❤2🤨2🤓1
“AI Anthropomorphization is bad bro.” Really? Explain this:
Biological study confirm that a new AI model’s behavior, created without biological inspiration, turns out to be a better model of the biological behavior than previous models too.
I.e. AI not being inspired by biology. AI being created first, and only later finding that biology does the same behavior, in a very different way.
Athropomorphization isn’t a fallacy, it’s more of the reality with each passing day in modern AI.
Many such cases.
2020 Nature Study by Deepmind
Blog Article
Biological study confirm that a new AI model’s behavior, created without biological inspiration, turns out to be a better model of the biological behavior than previous models too.
I.e. AI not being inspired by biology. AI being created first, and only later finding that biology does the same behavior, in a very different way.
Athropomorphization isn’t a fallacy, it’s more of the reality with each passing day in modern AI.
Many such cases.
2020 Nature Study by Deepmind
Blog Article
👍7👏3😐3😁1🤓1