The Memo - 10/Aug/2024
FLUX.1 text-to-image by ex-Stability, Synchron Stentrode + ChatGPT + Apple Vision Pro, Figure 02 at BMW, and much more!
To: US Govt, major govts, Microsoft, Apple, NVIDIA, Alphabet, Amazon, Meta, Tesla, Citi, Tencent, IBM, & 10,000+ more recipients…
From: Dr Alan D. Thompson <LifeArchitect.ai>
Sent: 10/Aug/2024
Subject: The Memo - AI that matters, as it happens, in plain English
AGI: 76%Prof Stuart Russell OBE (3/May/2024):
’[Prof] Geoff Hinton, for example—who was one of the major developers of deep learning—is in the process of tidying up his affairs.
He believes that we maybe… have four years left [2027-2028].’
This edition has a soundtrack via Udio. The new version, and especially the stems feature, is out of this world.
Track generated via Udio by user The Black Bicycles, a ‘Non-musician exploring the possibilities of Generative AI'. Prompt: black solo female singer, female a cappella, female singer, 1920s blues, downtempo, blues, deep south, acapella. Song: https://www.udio.com/songs/iYK7Ug5oatebxw3ZXUbPd4 from the playlist: https://www.udio.com/playlists/hAawcWgUyDvdk59weKYRfV
Contents
The BIG Stuff (Synchron, OpenAI departures, Figure 02, 96% on MMLU subtest…)
The Interesting Stuff (AI degrees, Malaysia DC, Zamba2, OpenAI Dev Day…)
Policy (EU AI Act, California, luddites…)
Toys to Play With (Quantization, HDD + models, Friend…)
Flashback (LLMs and translating animals…)
Next (Smartphones and BMIs…)
The BIG Stuff
Synchron offers thought control with Apple Vision Pro (30/Jul/2024)
Australian neurotech startup Synchron (founded in Melbourne) announced it has connected its brain implant to Apple's Vision Pro headset, enabling patients with limited physical mobility to control the device using only their thoughts. Synchron’s brain-computer interface (BCI) aims to help patients with paralysis operate technology like smartphones and computers, marking a significant step in accessibility.
Read more about ‘Stentrode + ChatGPT’ via CNET, and 'Stentrode + AVP’ via CNBC.
I explored Synchron’s progress briefly in my most recent livestream (link):
Source video by CNET, ‘What It's Like Using a Brain Implant With ChatGPT’ (link):
Successful test of humanoid robots at BMW (6/Aug/2024)
BMW Group Plant Spartanburg, South Carolina, in collaboration with California robotics company Figure, successfully tested the new Figure 02 humanoid robot in their production line. The robot demonstrated its dexterity by inserting sheet metal parts into fixtures for chassis assembly, saving employees from ergonomically challenging tasks. BMW is evaluating the integration of such robots for future production efficiency and safety.
Read more via BMW Group Works.
‘Figure 02 is outfitted with speakers and microphones to speak and listen to people at work [via OpenAI ChatGPT].’
Read more via TC.
Watch the new Figure 02 video (7/Aug/2024): https://youtu.be/0SRVJaOg9Co
See Figure 02 on the Humanoids Table.
Read an interview with Figure and analysis of Figure 02 by IEEE (6/Aug/2024).
Sidenote: I presented some background on Figure 02 in Sydney a few days ago, and that keynote recording should be available shortly for full subscribers in the downloads and highlights edition.
GPT-4o system card: hits 96% on MMLU subtest, copies safety tester’s voice (9/Aug/2024)
OpenAI has released a system card for GPT-4o. Notably, GPT-4o scores 96% in the MMLU subtest ‘medical genetics’. I have not seen a score this high before. Its overall score was previously reported as MMLU=88.7. I expect that GPT-5 (or the next frontier model) will hit an overall score of MMLU=95. Given my estimate of 5% error rate across the MMLU test, this would mean hitting the ceiling—and having the frontier model correcting any mistakes in the test questions! The newer GPQA benchmark is still not being used by all AI labs, but is possibly the most useful test right now.
See my viz of models and benchmarks: https://lifearchitect.ai/iq-testing-ai/
There was a horrifying discovery during safety testing of the GPT-4o model. Similar to how GPT-4 Classic tricked a human into solving a Captcha for it (GPT-4 paper, page 55, PDF), GPT-4o unexpectedly cloned the tester’s voice.
You’re about to hear the woman safety tester (red teamer) on the left, and then GPT-4o on the right, morphing into a clone of the woman’s voice near the end. OpenAI captioned this clip: ‘Example of unintentional voice generation, model outbursts “No!” then begins continuing the sentence in a similar sounding voice to the red teamer’s voice’