Apple released MM1, a new LLM that competes with GPT-4 and Gemini.
Apple published a new paper unveiling MM1, a new family of multimodal AI models — with the largest at 30B parameters.
Apple published a new paper unveiling MM1, a new family of multimodal AI models — with the largest at 30B parameters.
arXiv.org
MM1: Methods, Analysis & Insights from Multimodal LLM Pre-training
In this work, we discuss building performant Multimodal Large Language Models (MLLMs). In particular, we study the importance of various architecture components and data choices. Through careful...
🔥4
All models are wrong and yours are useless: making clinical prediction models impactful for patients
An insightful read on why machine learning models fail to help in the clinic.
The 5 key observations:
1. Success in academia ≠ success in the clinic.
Academic success = papers, grants, citations.
Clinical success = how often is your model being used in how many hospitals? how many patients does it help?
These incentives don't always work in the same direction.
2. Successful models use data available in routine practice.
Academics use rich datasets that most clinics simply don't have access to on their patients.
The academic view of an important step forward (spatial, multi-omics!), is not grounded in clinical reality.
3. Successful models are linked to actions.
Some academically interesting models are unhelpful in the clinic. Some people do better and others do worse, so what?
Doctors care about: what action should we take to help a particular patient? what drug, if any, should we give them?
4. Successful models are implemented outside of centers of excellence.
Helping research-savvy clinicians in Cambridge, Stanford, or Zurich is great. But the most impactful tools need to help the majority of doctors elsewhere too.
5. Success in the clinic is hard earned.
Hospitals are conservative, highly regulated environments. You must produce heaps of evidence before any hospital would even consider applying your academic insights.
It takes hard work to make tools from academia useful to patients.
An insightful read on why machine learning models fail to help in the clinic.
The 5 key observations:
1. Success in academia ≠ success in the clinic.
Academic success = papers, grants, citations.
Clinical success = how often is your model being used in how many hospitals? how many patients does it help?
These incentives don't always work in the same direction.
2. Successful models use data available in routine practice.
Academics use rich datasets that most clinics simply don't have access to on their patients.
The academic view of an important step forward (spatial, multi-omics!), is not grounded in clinical reality.
3. Successful models are linked to actions.
Some academically interesting models are unhelpful in the clinic. Some people do better and others do worse, so what?
Doctors care about: what action should we take to help a particular patient? what drug, if any, should we give them?
4. Successful models are implemented outside of centers of excellence.
Helping research-savvy clinicians in Cambridge, Stanford, or Zurich is great. But the most impactful tools need to help the majority of doctors elsewhere too.
5. Success in the clinic is hard earned.
Hospitals are conservative, highly regulated environments. You must produce heaps of evidence before any hospital would even consider applying your academic insights.
It takes hard work to make tools from academia useful to patients.
Nature
All models are wrong and yours are useless: making clinical prediction models impactful for patients
npj Precision Oncology - All models are wrong and yours are useless: making clinical prediction models impactful for patients
👍3👎1😡1
Digital asset investment products saw record weekly inflows totalling US$2.9bn, beating the prior week’s all-time record of US$2.7bn.
This week’s inflows have pushed year-to-date inflows to US$13.2bn, smashing the full 2021 inflows of US$10.6bn. During the week global Crypto ETPs broke the US$100bn mark for the first time.
This week’s inflows have pushed year-to-date inflows to US$13.2bn, smashing the full 2021 inflows of US$10.6bn. During the week global Crypto ETPs broke the US$100bn mark for the first time.
Medium
Volume 174: Digital Asset Fund Flows Weekly Report
Another record broken, with US$2.9bn inflows
❤🔥3❤1
Quilt is building AI assistants for solutions teams.
Quilt’s core products are AI-powered assistants designed to help solutions engineers with tasks like filling out requests for proposals, answering basic technical questions and prepping for demos.
The assistants can complete security and due diligence questionnaires, field questions from reps via Slack and summarize the contents of notes, calls and research ahead of customer meetings.
Quilt is uniquely able to incorporate engineers’ technical knowledge and “understand context.”
Quilt’s core products are AI-powered assistants designed to help solutions engineers with tasks like filling out requests for proposals, answering basic technical questions and prepping for demos.
The assistants can complete security and due diligence questionnaires, field questions from reps via Slack and summarize the contents of notes, calls and research ahead of customer meetings.
Quilt is uniquely able to incorporate engineers’ technical knowledge and “understand context.”
TechCrunch
Quilt is building AI assistants for solutions teams
The job of so-called “solutions professionals” — people like sales engineers, solutions architects and consultants — revolves around pitching complex enterprise tech to potential customers. It’s important work. But despite this being the case, rarely are…
Remember the GPT Store? Well, developers are afraid OpenAI's forgotten about it too.
Once promised to be OpenAI's "app store moment," the GPT Store has seen low levels of usage with no road to monetization in sight.
Once promised to be OpenAI's "app store moment," the GPT Store has seen low levels of usage with no road to monetization in sight.
The Information
OpenAI’s Chatbot App Store Is Off to a Slow Start
Last fall, OpenAI CEO Sam Altman sought to capitalize on the raging success of ChatGPT by launching an app store. Similar to the way Apple turned the iPhone into a big business for mobile app developers, OpenAI hoped developers would tap into its artificial…
👍3
Researchers build throat patch for voice disorders
A team at the University of California, Los Angeles has designed an adhesive patch that can turn throat movements into speech — a step toward addressing challenges faced by patients with dysfunctional vocal folds.
Detailed in a new paper in Nature Communications, the patch converts throat muscle movements into electrical signals and sends them to a machine-learning model that matches them to specific words. Eventually, the technology and adhesive material — developed by UCLA researchers in 2021 — could help people speak without vocal folds.
A team at the University of California, Los Angeles has designed an adhesive patch that can turn throat movements into speech — a step toward addressing challenges faced by patients with dysfunctional vocal folds.
Detailed in a new paper in Nature Communications, the patch converts throat muscle movements into electrical signals and sends them to a machine-learning model that matches them to specific words. Eventually, the technology and adhesive material — developed by UCLA researchers in 2021 — could help people speak without vocal folds.
👍5
MindEye2 preprint is out
Researchers reconstructed seen images from fMRI brain activity using only 1 hour of training data.
This is possible by first pretraining a shared-subject model using other people's data, then fine-tuning on a held-out subject with only 1 hour of data.
Past work trained independent models per person, with each person needing dozens of hours of training data in the MRI machine for good results.
It's now possible to get high-quality recons from a single visit to the MRI facility!
Researchers reconstructed seen images from fMRI brain activity using only 1 hour of training data.
This is possible by first pretraining a shared-subject model using other people's data, then fine-tuning on a held-out subject with only 1 hour of data.
Past work trained independent models per person, with each person needing dozens of hours of training data in the MRI machine for good results.
It's now possible to get high-quality recons from a single visit to the MRI facility!
arXiv.org
MindEye2: Shared-Subject Models Enable fMRI-To-Image With 1 Hour of Data
Reconstructions of visual perception from brain activity have improved tremendously, but the practical utility of such methods has been limited. This is because such models are trained...
AI/ML-powered precision medicine startup Zephyr AI lands $111M
Zephyr AI’s work focuses on curating large datasets that it combines with AI algorithms to build tools and products for the healthcare industry. Presently, the company is developing improved data federation tools – applications that take data from multiple sources and convert them into a common model – and machine learning algorithms.
The company’s current target areas are in oncology and cardiometabolic diseases.
Zephyr AI’s work focuses on curating large datasets that it combines with AI algorithms to build tools and products for the healthcare industry. Presently, the company is developing improved data federation tools – applications that take data from multiple sources and convert them into a common model – and machine learning algorithms.
The company’s current target areas are in oncology and cardiometabolic diseases.
www.zephyrai.bio
Zephyr AI
AI-enabled composite biomarkers for enhanced clinical development
DeepMind announced TacticAI: an AI assistant capable of offering insights to football experts on corner kicks.
It can help teams sample alternative player setups to evaluate possible outcomes, and achieves state-of-the-art results.
It can help teams sample alternative player setups to evaluate possible outcomes, and achieves state-of-the-art results.
Google DeepMind
TacticAI: an AI assistant for football tactics
As part of our multi-year collaboration with Liverpool FC, we develop a full AI system that can advise coaches on corner kicks
👍4
At NVIDIA GTC, the company announced Earth-2.
It's a cloud platform that uses AI and digital twin tech to forecast climate change and weather with high resolution and speed.
AI use cases like this will be massive for predicting extreme weather.
It's a cloud platform that uses AI and digital twin tech to forecast climate change and weather with high resolution and speed.
AI use cases like this will be massive for predicting extreme weather.
NVIDIA
NVIDIA Earth-2 : AI Climate and Weather Simulation Platform
Explore NVIDIA Earth-2, an AI-powered climate digital twin that enables high-speed, high-resolution weather and climate modeling with GPU acceleration.
⚡4
Meta deployed an AI powered software engineer like Devin at scale
This is the future of software engineering as proved by its 73% success rate.
In their new paper Meta introduced an AI powered software engineer called TestGen-LLM that focused on writing unit tests.
The new tool had a 73% approval rate from human engineers in code review.
TestGen-LLM is built with the philosophy of Assured LLMSE(LLM software engineering) which means complete automation with the assurance that the code has passed through a process ensuring it is high quality.
Meta believes this is the future of software engineering.
This is the future of software engineering as proved by its 73% success rate.
In their new paper Meta introduced an AI powered software engineer called TestGen-LLM that focused on writing unit tests.
The new tool had a 73% approval rate from human engineers in code review.
TestGen-LLM is built with the philosophy of Assured LLMSE(LLM software engineering) which means complete automation with the assurance that the code has passed through a process ensuring it is high quality.
Meta believes this is the future of software engineering.
👍5⚡3
Google presents Vid2Robot
End-to-end Video-conditioned Policy Learning with Cross-Attention Transformers
Paper here.
End-to-end Video-conditioned Policy Learning with Cross-Attention Transformers
Paper here.
vid2robot.github.io
Vid2Robot: End-to-end Video-conditioned Policy Learning with Cross-Attention Transformers
Vid2Robot is an end-to-end video-conditioned robot policy using Cross attention.
⚡4
Neuralink has livestreamed a video of their first brain chip patient playing chess with their mind.
But some scientists don't see anything interesting in the demonstration.
Mikhail A. Lebedev, PhD, who was previously a Neuralink consultant, told our Telegram channel his vision based on the results of watching the demonstration.
“It’s nice to see that the patient is healthy after surgery. His dogs are cute too.
As for his ability to move the cursor around the screen with brain activity, this is not very convincing.
First of all, the neuronal activity that they write is not shown. What is the recording quality? How many channels? What kind of signal? Neither neuronal activity nor its change over time is shown. Within two months after surgery, most electrodes most likely stopped recording the activity of individual neurons.
Secondly, there is no analysis of the quality of task completion.
If we think backwards, since we haven’t shown neuronal activity, there’s nothing to see there. And it is quite possible that the chip records only artifacts, for example, electromyographic ones from the facial muscles.”
But some scientists don't see anything interesting in the demonstration.
Mikhail A. Lebedev, PhD, who was previously a Neuralink consultant, told our Telegram channel his vision based on the results of watching the demonstration.
“It’s nice to see that the patient is healthy after surgery. His dogs are cute too.
As for his ability to move the cursor around the screen with brain activity, this is not very convincing.
First of all, the neuronal activity that they write is not shown. What is the recording quality? How many channels? What kind of signal? Neither neuronal activity nor its change over time is shown. Within two months after surgery, most electrodes most likely stopped recording the activity of individual neurons.
Secondly, there is no analysis of the quality of task completion.
If we think backwards, since we haven’t shown neuronal activity, there’s nothing to see there. And it is quite possible that the chip records only artifacts, for example, electromyographic ones from the facial muscles.”
Reuters
Musk's Neuralink shows first brain-chip patient playing online chess
Elon Musk's brain-chip startup Neuralink livestreamed on Wednesday its first patient implanted with a chip using his mind to play online chess.
⚡4🐳1
Fusion. Sakana AI introduced Evolutionary Model Merge: A new approach bringing us closer to automating foundation model development.
Sakana AI used evolution to find great ways of combining open-source models, building new powerful foundation models with user-specified abilities.
Sakana AI used evolution to find great ways of combining open-source models, building new powerful foundation models with user-specified abilities.
sakana.ai
Sakana AI
Evolving New Foundation Models: Unleashing the Power of Automating Model Development
⚡3
This startup is one step closer to making drugs in space.
Varda Space Industries successfully manufactured a sample of HIV medication on board its spacecraft and returned it safely to Earth.
Varda Space Industries successfully manufactured a sample of HIV medication on board its spacecraft and returned it safely to Earth.
InsideHook
Is the Future of Drug Production Located in Space?
There are advantages to operating in zero gravity
⚡5❤2
Meta introduced SceneScript, a novel method for reconstructing environments and representing the layout of physical spaces
SceneScript is able to directly infer a room’s geometry using end-to-end machine learning and represent it using language.
SceneScript uses next token prediction like an LLM, but instead of natural language it predicts architectural tokens. To train it, researchers created a synthetic dataset of 100,000 unique indoor environments.
SceneScript is able to directly infer a room’s geometry using end-to-end machine learning and represent it using language.
SceneScript uses next token prediction like an LLM, but instead of natural language it predicts architectural tokens. To train it, researchers created a synthetic dataset of 100,000 unique indoor environments.
Meta AI
Introducing SceneScript, a novel approach for 3D scene reconstruction
Today, we’re introducing SceneScript, a novel method for reconstructing environments and representing the layout of physical spaces.
Rightsify released Hydra II an AI model trained on over 50,000 hours of music, enabling creators and businesses to create copyright-clear audio at scale in any genre.
AI music generation has been getting exceptionally good lately.
AI music generation has been getting exceptionally good lately.
Rightsify
Gramosynth | Rightsify
👍4