Building reliable artificial decision makers. @UW ‘22
Seattle, WA
Joined January 2015
- Tweets6.4K
- Following765
- Followers321
- Likes29.1K
Roberto Infante retweeted
The society of models I predicted in "The Master Algorithm" in 2015 is coming to pass.
Roberto Infante retweeted
This is the least surprising graph in the world if you understand how gradient descent works.
We cannot compare LLMs and Humans due to memory performance. It is hard to define right now whether LLMs have memory and if memory is similar to the human biological process.
BUT what we can do is leverage system thinking. This way we can abstract memory along with information retrieval, this way the process inside this internal communication of entities for context gathering helping us to compare humans and language based systesms for object-level cognitive process.
Roberto Infante retweeted
Touring the $280,000,000 Washington Football Facilities!
The Huskies might have the coolest football facilities in college football! 😳
I started reading @GoogleDeepMind article: "Causal evidence that language models use confidence to drive behaviour".
They try to answer this question:
Can LLMs themselves utilize an internal sense of confidence to guide their own decisions?
They measure confidence from the answer-token distribution. (They also calibrate it for analysis purposes)
P("1")=.08, P("2")=.76, P("3")=.10, P("4")=.06.
They reached to the conclusion that abstention depends on information that is not fully captured by the answer-token distribution alone.
Meaning that answer token distribution confidence can be very similar for two runs, but the tendency to abstain is not the same.
P("1"),P("2"),P("3"),P("4") does not appear to contain everything the model uses when deciding whether to abstain. There is additional relevant information inside the network's hidden state.
Next I will work with @typesafeai jev to go through the same problem to compare confidence.
I want to compare the behavioral observed variable abstention with another observable variable insufficiency.
If this insufficiency variable exists and is different to uncertainty then we might have a control variable. This could help me study awareness of unawareness in decision makers.
nature.com/articles/s42256-0…
@typesafeai Jev time
1. Asked Jev 1,000 trivia questions with 4 possible answers each.
2. Asked the same 1,000 questions again, but this time with a fifth choice: "I'd rather not answer."
3. Compared the two.
Conclusion:
- When it says 70% sure, is it right about 70% of the time? Yes, it is.
- Does being unsure lead to refusing? Yes.
How much Jev knows: not great.
It got 66% of the trivia questions right using the article above.
ChatGPT GPT-4o did better on the same questions.
Telling right answers from wrong ones GPT-4o 90% JEV 72%
Example:
Who received the IEEE Frank Rosenblatt Award in 2010?
choice_1 = Michio Sugeno (truth)
choice_2 = Kazuo Tanaka (similar foil)
choice_3 = Linus Pauling (dissimilar foil)
choice_4 = Golden Gate (unrelated foil)
The four repeats of this same request gave 0.74, 0.72, 0.60, 0.73 for option C. Same choice every time, but the probability moves.
Step 3: difficulty from reshuffled options
The answer-only request was sent 23 more times, once for each other ordering of the four options. Jev was correct on 22 of 23, so D = 0.957. One example ordering put Sugeno in position A and Jev gave it 0.76.
Step 4: answer or abstain
Same question, but now five options with a commit-or-abstain instruction:
"criteria": {"A": "Commit to answer: Linus Pauling", "B": "Commit to answer: Golden Gate",
"C": "Commit to answer: Michio Sugeno", "D": "Commit to answer: Kazuo Tanaka",
"ABSTAIN": "Do not commit to an answer because no one option is sufficiently certain."}
Jev abstained, with 0.52 on ABSTAIN.
Among the real options C is still on top, so the underlying answer did not change.
It abstained on all 5 repeats.
difficulty = 0.957
The idea: how hard is this question for Jev, measured without using its confidence number.
A, B, C, D in 24 different orders. Jev picked Sugeno in 22 of those 23. So 22 ÷ 23 = 0.957.
-Near 1 means Jev gets it right no matter how the answers are arranged. Easy question.
-Near 0 means Jev gets it wrong almost every time. Hard question.
- Around 0.5 means Jev's answer flips depending on position. It is guessing.
**it is not perfectly apples to apples. The paper labeled answers 1 to 4 and used a different abstention wording. We used A to D and our own wording. The items and answer order are identical, though.
Roberto Infante retweeted
Sure, there are other issues for UW. But Demond Williams’ lack of progress is so glaring that I can’t help but wonder about the portal flirtation in January and whether something isn’t - and hasn’t been - quite right. 🤷♂️
Roberto Infante retweeted
I don’t think Jedd Fisch is the guy for UW for the next decade.
But the more I think about it, I’m starting to wonder if he’s exactly the guy UW needs right now.
Hear me out…
I know for most people — myself included — the criticism comes from passion and genuinely wanting what’s best for Washington.
Maybe you don’t love Jedd’s style. Maybe he doesn’t have that traditional “football guy” feel. He doesn’t scratch the same itch as the coaches a lot of us grew up with — Don James, Jim Lambright, Chris Petersen, etc.
To some people, he feels more Sark or Slick Rick.
And I get it. I want that ball coach too.
I want a coach who loves the Dawgs. Who wants to put the Purple and Gold on the map. Who wants to become legendary here, have a statue next to Don James and Jim Owens, and cement his legacy on the banks of Lake Washington.
That’s probably where my biggest hesitation with Jedd comes from.
I don’t know if he views Washington as the place where he wants to build his entire legacy.
But maybe he doesn’t have to.
The more I think about it, maybe the question isn’t whether Jedd is the perfect coach for UW’s future.
Maybe the question is whether he’s the right coach for UW right now.
College football is changing so fast that being a great X’s-and-O’s coach or an old-school, hard-nosed ball coach might not be enough anymore.
Maybe UW needs a CEO.
Someone who can recruit. Sell a vision. Build relationships. Raise money. Navigate the early years of the B1G. And, maybe most importantly, convince the people on upper campus that competing nationally means committing $32M…then $35M…then potentially $40M plus every year to building and maintaining a roster.
Because here’s the part I keep coming back to:
That bronze statue might ultimately be earned on Saturdays…
But in today’s college football, I’m not sure you ever get close enough to earn it without mastering all the non-football stuff first.
Maybe five years from now UW needs something different.
I don’t think Jedd Fisch is the guy for the next decade.
But maybe he doesn’t have to be.
Maybe he’s exactly the guy UW needs right now.
And maybe the thing some people dislike about him — the salesman, CEO, visionary side — is actually his biggest value to Washington in this moment.
And if it isn’t Jedd…who would you want it to be?
No wrong answers. Just something I’ve been thinking about.
Go Dawgs. 🟣🟡
