The BIG NEW THING in AI research. Because normal results for seq2sseq is already doing very well. Slightly more realistic models of language acquisition.

“Linguistics is not just about human languages”: humans, animals, and machines.

BIG AGI What: question is the difference between a dumb human and GPT?

Speech is Based for Interoperability

• continuous data (as opposed to text)
• much less complex with vision: vision is more complex
• speech is a nice controllable system (????)

Models of Language Acquisition

https://arxiv.org/pdf/2309.07861.pdf

A lecture hosted by Cynthia Lee.

“AI: how it works & why its often so biased”

Defining Artificial Intelligence.

AI/Clinical Decision Support

• workload measurement
• clinical wellness, etc.
• perioperative outcomes

Big problem: integration is impossible; there’s lots of models. “Researched models are rarely implemented; implemented models are rarely researched”. Epic doesn’t stand behind its models.

Implementation of AI based mechanisms though “saves time” on paper, results in more patient throughput and patient burn. At this point: harder question—not whether you can make a model, but how do you govren their use and actually put them into implementation.

Safety as a property is dependent on the being used. Meaning: “AI safety focuses on technical solutions to ensure that AI systems operate safely and reliably.”

• preventing accidents, misuse, and harmful consiquences
• machine ethics and AI alignment
• monitoring systems for risks
• developing norms and policies that promote safety

Logistics

Lectures

• SU-CS120 SEP242024
• SU-CS120 OCT082024
• SU-CS12 0OCT102024

AIBridge is an introductory AI bootcamp developed and taught by Prof. Xin Liu, yours truly, and Samuel Ren in collaboration with AIFS.

AIBRidge Notes

• Pause [more] to allow some time to see if people follow
• did y’all not introduce pandas?

Closest to doing this without try/except:

• slide 49: what is conc?
• is this too much recap time? Haven’t we been recapping for a long while already?
• probably good to mention what is /content/iris.data, also, just opening from ./iris.data should work and will be probably more ergonomic
• read function confusion
  • .read() => str
  • .readlines() => [str]
  • the pauses feel a tad ackward?
  • speak up!

• SSE squares and lines need to be darker: increase opacity 39
• “very common metric” — not a metric
• motivate confidence value better; the “middle” question makes sense
• I think its actually probably good to explain cross-entropy in the future
  • (i.e. its not a lot of fancy math + I think it provides a lot of intuition w.r.t. one-hot encoding, probablitiy distributions, etc.)
• Problem with how I made the old slides: multi-Class classification (1va, ava, etc.) needs better motivation before, otherwise throwing three classes on the screen is a tad confusing
• motivate that the whole random.seed business is so that the whole class can compare answers more effectively
• LogReg = LogisticRegression(), typically, name instance variables as lower snake case; so maybe call it my_log_reg or something