On May 16th, 2023 Sam Altman of OpenAI; Gary Marcus, professor at New York University, and Christina Montgomery, chief privacy and trust officer at IBM spoke to congress on topics related to AI regulation. A link to the hearing can be found here: Youtube: CNBC Senate hearing on AI oversight....
I am compiling a list of tasks and evaluations that are used to test LLMs. I intend to expand this list to include the initial published date, scope of questions, number of questions, direct links to the data-sets, question types (ie, multiple choice or fill-in the missing word, etc), along...
On May 16th, 2023 Sam Altman of OpenAI; Gary Marcus, professor at New York University, and Christina Montgomery, chief privacy and trust officer at IBM spoke to congress on topics related to AI regulation. A link to the hearing can be found here: Youtube: CNBC Senate hearing on AI oversight.
From a lens of AI Alignment, the general substance of the conversation focused on near term effects such as job loss, bias, harmful content, targeted advertising, privacy implications, election interference, IP and copyright issues and other similar topics. Sam Altman has spoken about hard AI risks before, but he was not explicit about them in the hearing. Gary Marcus communicated that his estimation... (read 386 more words →)
I am compiling a list of tasks and evaluations that are used to test LLMs. I intend to expand this list to include the initial published date, scope of questions, number of questions, direct links to the data-sets, question types (ie, multiple choice or fill-in the missing word, etc), along with additional comments on perceived difficulty and other characteristics. The majority of automated test suites rely on multiple choice answer prompts, as open question free-form questionnaires are difficult to evaluate.
TruthfulQA: https://github.com/sylinrl/TruthfulQA
MMLU: https://github.com/hendrycks/test
HellaSwag: https://github.com/rowanz/hellaswag/tree/master/data
WinoGrande: https://github.com/allenai/winogrande
HumanEval: https://github.com/openai/human-eval
DROP: https://arxiv.org/abs/1903.00161
GSM8K: https://github.com/openai/grade-school-math
LogiQA: https://github.com/lgw863/LogiQA-dataset
CoQA: https://stanfordnlp.github.io/coqa/
LAMBADA: https://zenodo.org/record/2630551#.X4Xzn5NKjUI
ReClor: https://whyu.me/reclor/
BoolQ: https://arxiv.org/abs/1905.10044
PIQA: https://yonatanbisk.com/piqa/
SIQA: https://arxiv.org/abs/1904.09728
AI2 Reasoning Challenge (ARC) 2018: https://allenai.org/data/arc
RACE: https://www.cs.cmu.edu/~glai1/data/race/
More tests are enumerated on page 27 in this paper: more on "A... (read more)
I looked over a bit of David's public facing work, eg: https://www.youtube.com/watch?v=I7hJggz41oU
I think there is a fundamental difference between robust, security minded alignment and tweaking smaller language models to produce output that "looks" correct. It seems David is very optimistic about how easy these problems are to solve.
It seems one could convince this hypothetical emperor to invest into industrialization of technology by offering to build things other than a steam engine, or outlining how a steam engine leads to them -- telegraph, or semaphore towers to send news of invasions or changes in distant towns or provinces, better manufacturing capability for tools and weapons, food storage and transport mechanisms, etc.
I wonder how much the survey paying a few RMB and asking for a wechat ID influenced the results. Great work working and sharing this poll though!
While this survey’s responses are anonymous, respondents did submit their WeChat IDs so that they could be remunerated for participation.
The current link to the podcast is available here: https://munkdebates.com/podcast/the-rise-of-thinking-machines
Top 10 donations in 2023, since the html page offers no sorting and is sorted by date:
| $2,800,000 | Cooperative AI Foundation | General support |
| $1,846,000 | Alignment Research Center | General support for ARC Evals Team |
| $1,733,000 | Center for Applied Rationality | General support for Lightcone Infrastructure |
| $1,327,000 | Center on Long-Term Risk | General support |
| $1,241,000 | Manifold for Charity | General support for Manifold Markets |
| $1,159,000 | Alliance to Feed the Earth in Disasters | General support |
| $1,000,000 | Carnegie Mellon University | Foundations of Cooperative AI Lab |
| $1,000,000 | Massachusetts Institute of Tech | Gift to the Tegmark research group at MIT for General Support |
| $1,000,000 | Meridian Prime | General support |
| $909,000 | Center for Artificial Intelligence Safety | General support |
A bit of feedback: the "We get a second chance at building AGI" outcome should not be an outcome or perhaps rephrased.
I tracked down the exact quote where Prof Marcus was talking about timelines with regards to jobs. He mentioned 20-100 years (right before the timestamp) and then went on to say: https://youtu.be/TO0J2Yw7usM?t=2438
"In the long run, so called AGI really will replace a large fraction of human jobs. We're not that close to AGI, despite all the media hype and so forth ... in 20 years people will laugh at this ... but when we get to AGI, let's say it is 50 years that is really going to have profound effects on labor..."
Christina Montgomery is explicitly asked "Should we have one" [referring to a new agency] by Senator Lindsey Graham and says "I don't think so" at https://youtu.be/TO0J2Yw7usM?t=4920
Couple of more takeaways I jotted down:
PaLM2 followed closely [to] Chinchilla optimal scaling. No explicit mention of number of parameters, data withheld. Claim performance is generally equivalent to GPT-4. Chain-of-thought reasoning is called out explicitly quite a bit.
Claims of longer context length, but no specific size in the technical report. From the api page: "75+ tokens per second and a context window of 8,000 tokens,"
"The largest model in the PaLM 2 family, PaLM 2-L, is significantly smaller than the largest PaLM model but uses more training compute" "The pre-training corpus is significantly larger than the corpus used to train PaLM [which was 780B tokens]"
I was somewhat disturbed by the enthusiastic audience applause to dire serious warnings. What are techniques or ways to anchor conversations like this to keep them more serious?
I thought the criticism on that specific quote was that the "higher intelligence" group, while taking more time, did solve the hard questions correctly, as opposed to not solving them correctly at all.