1 of 66

Generative AI�How it works & applications for media orgsTim Olson, KQED tolson@kqed.org

http://bit.ly/48ccxIo

2 of 66

AGENDA

How foundation models work

Challenges - bias and hallucinations

Changes - media business & tools

Examples - media orgs applications

Discussion - talk with each other, headlines

3 of 66

HOW FOUNDATION MODELS WORK

Set goal

Lots of�data

Neural network

Train

Fine-tune

4 of 66

Identify an object in a photo

Given a sequence of text, guess what comes next

Predict the three-dimensional shapes of proteins

Set goal

5 of 66

Wikipedia, Reddit, Twitter, books, manuals, comments...

Photos

Amino acid sequences

Lots of�data

> data = model improves

6 of 66

A complex web of interconnected nodes (or “neurons”) that process and store information

Neural network

> compute = model improves

7 of 66

Break words into tokens - basic units than can be encoded.

Produce a vector (aka word embedding) (list of number values) -�- hundreds of values, each representing a different aspect of a words meaning

Neural network

8 of 66

Transformer model of neural network can analyze multiple pieces of tokens at the same time�(Chat GPT = Generative Pretrained Transformer)

Neural network

9 of 66

As it analyzes the data, token by token, it identifies patterns and relationships

Train

learns from the data itself

10 of 66

Develops a sense of context, it can also pick up other, unexpected abilities, such as knowing how to code

Train

11 of 66

Reinforcement learning with human feedback

Fine-tune

12 of 66

Retrieval Augmented Generation

Grounding

survey

email

website

medical

financial

legal

foundational model +�specific data set

13 of 66

14 of 66

Challenges

Data source - e.g. Reddit, web, articles… develop bias due to untruths, hate speech ..

Generate wrong info (hallucination) - generate (untruth) text not in the training data. Unlike classical computing where given the same inputs, you get the same output, Generative AI systems spin out multiple possibilities from a single prompt (given trillions of variables)

15 of 66

Generative is� non deterministic� (not traditional programming, results are probabilistic)�

Good for Poor for

Explore possibilities Reliable answer finder

Ideation Determining truth

Code development Decision maker

Data analysis A replacement for creativity

via: Toshi Anders Hoo, Institute for the Future

16 of 66

Cautions

AI makes mistakes

Be transparent about use of AI

Do not upload sensitive information e.g. proprietary data

Do not publish any AI generated text without human editorial

�Appropriate: Data analysis, repetitive tasks, process improvement, first drafts, generating ideas

Not appropriate: Ethical concerns (e.g. promotion decisions), legal risks, privacy and security violation, discrimination risk, replace human interaction (e.g. difficult conversation)�

KQED AI guidelines

Nebraska Public Media Framework for Generative AI Experimentation

Principles for Using Generative A․I․ in The Times’s Newsroom | The New York Times Company

17 of 66

AI - THE NEXT BIG SHIFT

PC & Servers

Web & Internet

Cloud & Mobile

AI

Satya Nadella, Microsoft CEO - My annual letter: Leading in a new era

18 of 66

AI - THE NEXT BIG SHIFT

natural language (text)

vision

audio

foundation model

completing, summarizing, detecting anomalies, identify patterns, surface insights…

data

traffic, ocean currents, amino-acids, crop disease, radiology, law, stocks, star light spectrum…

19 of 66

vision

bird classifications

20 of 66

vision

blight

weed

soil

21 of 66

vision

tissue type

pneumothorax

skeletal anomaly

22 of 66

language

audio

Wendy’s menu

23 of 66

language

audio

crow vocal repertoires

24 of 66

trained on historical data to learn complex systems and generate projections, whereas numerical models use physics equations — and weather observations — to produce simulations of future conditions

GraphCast

weather

25 of 66

RFM-1

physics

space

measurement

language

audio

Physical AI�perceive, reason and act in�3D space and time

26 of 66

discovered more than 2.2 million hypothetical materials, including 381,000 stable new materials

microchips

batteries

photovoltaics

GNoME

inorganic crystals

27 of 66

generate biomolecular structure predictions containing proteins, DNA, RNA, ligands, ions

AlphaFold

protein structure

drug development�from discovery to engineering

28 of 66

TECH INVESTS TO BE THE AI “PLATFORM”

natural language (text)

vision

audio

foundation model

completing, summarizing, detecting anomalies, identify patterns, surface insights..

data

traffic, amino-acids, crop disease, radiology, law, stocks, star light spectrum

agent

agent

agent

agent

agent

29 of 66

ERNIE 4.0

GPT-4.o

Gemini

LLaMa-3.2

Mistral Large

Claude 3

Apple Intelligence

Invests in Anthropic

30 of 66

DOWNSIDES

Misinformation- deep fakes, people stop trusting true information, the “liar’s dividend”

Security �- cyber attack, spoofing, microtargeting phishing…

Energy consumption�- training models and queries take more energy; though smaller models are being developed too

Jobs lost, added, & changed

31 of 66

Change is accelerating�

How do we help society understand and adapt?

How do we make the positive changes more equitable?

What will the market not address that society needs?

32 of 66

MEDIA�BUSINESS

33 of 66

AI driven intelligent devices

language

audio

vision

34 of 66

Transcription, translation, dubbing

Translation tipping point - Apple 17.4. podcast transcription, Samsung Galaxy S24 live translation

Localize to multiple markets - translation, captoning, subtitling, dubbing, lip synch..

language

audio

35 of 66

Personalized advertising

6,000 taglines

Project brief

Brand archetypes

Example taglines

location

weather

36 of 66

language

audio

character

37 of 66

companion

language

audio

38 of 66

Search

“Zero-click”

Links→Answers

Threaded

Conversational

language

39 of 66

40 of 66

Summarization� at scale and personalized

AI chips on mobile device summarize news + for user

�Your “assistant” for calendar, events, news, traffic, weather

41 of 66

Sign

Sue

or

42 of 66

Grounding

Opinion | Ex-Google director: The real wolf menacing the news business is AI - The Washington PostJim Albrecht, senior director of news ecosystem products at Google from 2017 to 2023

“generative AI products tend to rely on a process known as grounding,” in which the statements made by the AI are checked against relevant source documents to ensure that the AI is not making things up. This process is especially critical if a user is asking about a recent event in which the relevant facts did not exist at the time of the LLM’s training. In such cases, the AI can only answer accurately if it retrieves those facts from recent grounding documents. These documents are the essence of the work newspapers do — sourcing and reporting new facts — and the fruits of that labor should reasonably belong to those who perform it

43 of 66

MEDIA�PRODUCTION

44 of 66

AI MEDIA PRODUCTION TOOLS

TEXT

IMAGE

  • ChatGPT OpenAI
  • Copilot Microsoft
  • Gemini Google
  • Claude.ai Anthropic
  • ChatGPT OpenAI
  • Copilot Microsoft
  • Gemini Google
  • Midjourney
  • Firefly Adobe

AUDIO

VIDEO

  • Sora OpenAI
  • Runway
  • Opus Clip
  • Pika Labs

45 of 66

46 of 66

47 of 66

AI - YOUR ASSISTANT (co-pilot, creative assistant)

48 of 66

AI - YOUR ASSISTANT (co-pilot, creative assistant)

high school intern

college grad

postgrad grad

PH.D

?

49 of 66

Lead with our values

People first

  • act in the best interests of the public
  • prioritize talent and creativity
  • open and transparent

BBC AI Principles

50 of 66

WHAT SHOULD WE DO

Learn & play - understand and help our community understand, try it

Processes - how can this save our staff’s time, reduce mundane tasks?

Explore - how we might leverage the tools for public service?

51 of 66

LEARN

GENERAL

JOURNALISM

52 of 66

PLAY

labs.google.com/search

search

personalize chatbot

homework

gramar

lensa - images

research documents

images

53 of 66

PROCESS

Generate draft text

  • Article headline brainstorm
  • Personalizing donor comms
  • Marketing copy in brand voice

Analyse a document

  • Transcript - identify speakers
  • Identify prospects in public docs
  • Contract - review terms

Audio workflow

  • Scratch track for time

Video workflow

  • Generate SEO keywords
  • Edit longform video to a short

54 of 66

EXPLORE �EXAMPLES OF PUBLIC FACING APPLICATIONS IN JOURNALISM

55 of 66

Context aware search of audio archives

How KQED is enriching its 'Forum' archive with generative AI - Current KQED

56 of 66

Audio�transcription�

  • themes
  • speaker identification
  • quotes
  • summary

57 of 66

Investigative journalism

(perception)

58 of 66

Investigative journalism

The California Reporting Project ��Law enforcement documents - 7.5 TB of data (written, text, body cam video..)

Newsgathering collaboration - request and analyze records from law enforcement agencies to report on use of force and misconduct cases around the state

OCR

LLM C�check LLM B

Toss all data, keep structure & redo

Missing data

LLM A

LLM B

59 of 66

Image generation�Photoshop Generative Expand

Filter�Photoshop

Style�After Effects

Branding & color�After Effects

60 of 66

Google Notebook LM

https://notebooklm.google.com/ ��

61 of 66

ChatBots based on publisher data

62 of 66

Content into other formats

NY Times Audio- automated voicing of articles

Opus Clip - vertical video shorts from long form horizontal video

63 of 66

Access, analysis and visualization of public data

Digital Democracy CalMatters

64 of 66

Access, analysis and visualization of public data

MinutesMichigan Public Radio

65 of 66

Access, analysis and visualization of public data

Amazon Mining Watch

Pulitzer Center´s Rainforest Investigations Network and Earthrise Media �

66 of 66

DISCUSSION

What are you seeing in your area?��How are you using it in your work?

http://bit.ly/48ccxIo

Tim Olson tolson@kqed.org