Upgrade to Pro — share decks privately, control downloads, hide ads and more …

The Rise of AI: Peak Ace Dinner Series (Munich)

The Rise of AI: Peak Ace Dinner Series (Munich)

Some thoughts on the current state of AI, presented at the Peak Ace Pre-SMX Munich Dinner Event 2023.

Bastian Grimm
PRO

March 23, 2023
Tweet

More Decks by Bastian Grimm

Other Decks in Marketing & SEO

Transcript

  1. Bastian Grimm | Peak Ace AG | @basgr
    The Rise of AI
    A few thoughts to chew on

    View Slide

  2. There is soooooo much going
    on right now…

    View Slide

  3. pa.ag
    @peakaceag
    3
    A timeline of AI and
    language models
    Source: https://pa.ag/3Ld7yxv

    View Slide

  4. pa.ag
    @peakaceag
    4
    GPT-4 will be much more extensive than its predecessor
    Attention - speculation: The parameter scope of GPT-4 is not confirmed yet and is only
    based on an interview with Sam Altman [CEO OpenAI].
    Source: https://pa.ag/3UgkBjm
    billion
    parameters
    1.5
    GPT-2
    billion
    parameters
    175
    GPT-3
    trillion
    parameters
    100
    GPT-4
    x571
    x177

    View Slide

  5. pa.ag
    @peakaceag
    5
    Problems with (large) language models
    ”Training a single BERT base model (without parameter tuning) on GPUs was estimated
    to require as much energy as a trans-American flight.” [Source: Mozilla]
    Source: https://pa.ag/3tuOwJ9
    Computing power
    The quality of the results
    improves as the size of the
    LLMs increases. Gains are
    incremental, while costs
    increase exponentially.
    Biased training data
    If biased data are used, the
    system will likely show the
    same bias when making
    decisions in practice.
    Environmental impact
    Training a large transformer
    model produces 5.680%
    more carbon emissions than
    one human per year.
    Static perspectives
    Large language models
    are usually trained once
    and used over long
    periods of time.

    View Slide

  6. pa.ag
    @peakaceag
    6
    A more hands-on “bias example” from Amazon
    The company developed AI to help scale recruiting – which didn’t work so well…
    Source: https://pa.ag/4228GtP
    Amazon’s computer models were trained
    to vet applicants by observing patterns in
    resumes submitted to the company over a
    10-year period. Most came from men,
    a reflection of male dominance across the
    tech industry.
    […]
    In effect, Amazon’s system taught itself
    that male candidates were preferable.
    It penalised resumes that included the
    word “women’s” and downgraded
    graduates of two all-women’s colleges.

    View Slide

  7. What took five years will not take five years again
    Unparalleled acceleration

    View Slide

  8. New tech = new challenges

    View Slide

  9. pa.ag
    @peakaceag
    9
    Ever wondered how those “make person XYZ sing a song”
    apps work?
    Source: https://pa.ag/3ZXB5PW
    The LSN library
    implementation
    is available for
    free on Github.

    View Slide

  10. pa.ag
    @peakaceag
    10
    Have you ever heard of DeepFaceLab by chance?
    Source: https://pa.ag/3yslyfn
    De-age the face
    Replace the head
    Replace the face

    View Slide

  11. pa.ag
    @peakaceag
    11
    DeepFaceLab: a deepfake framework for face-swapping
    It provides tools and an easy-to-use way to conduct high-quality face-swapping; and
    yes, it’s free
    Source: https://pa.ag/3YG9MbQ
    Deepfake defense not only requires
    the research of detection but also
    requires the efforts of generation
    methods. However, current deepfake
    methods suffer the effects of obscure
    workflow and poor performance. To
    solve this problem, we present
    DeepFaceLab […]

    View Slide

  12. pa.ag
    @peakaceag
    12
    DeepFaceLive: yep, you can even do this “live”…
    You can swap your face from a webcam or the face in any other video with trained face
    models which are ready to use out of the box:
    Source: https://pa.ag/3J9h9mo
    These persons
    do not exist.
    Similarities with
    real people are
    accidental.

    View Slide

  13. Stock Photo or AI?
    Which image was created by a machine?

    View Slide

  14. (no) Rules
    #A or #B – which image was generated by AI?

    View Slide

  15. VS
    #A #B

    View Slide

  16. VS
    #A #B

    View Slide

  17. VS
    #A #B

    View Slide

  18. VS
    #A #B

    View Slide

  19. A Some ideas to
    increase efficiency.

    View Slide

  20. pa.ag
    @peakaceag
    20
    The 2023 MAD (Machine Learning, Artificial Intelligence
    & Data) landscape: 1,400+ and growing
    Source: https://pa.ag/3LdKDST

    View Slide

  21. pa.ag
    @peakaceag
    21
    Source: https://pa.ag/3yrqQYn
    theresanaiforthat.com
    – yup… there is!
    If you prefer searching, this
    might be a good help:

    View Slide

  22. That‘s it; job done!?
    #kthxbye

    View Slide

  23. pa.ag
    @peakaceag
    23
    humata.ai: Summarise long papers (aka PDFs)
    You’re in SEO and reading, things like patents and other papers? Here’s your solution:
    Sources: https://pa.ag/3YyKhJv & https://pa.ag/3l6o0VR
    Or you can try chatpdf.com:

    View Slide

  24. pa.ag
    @peakaceag
    24
    Want to get more hands-on? Try Luminar AI photo editor
    Easy-to-use, available on any device (Win, Mac, Photoshop plug-in, etc)
    Source: https://pa.ag/3yuYxIA

    View Slide

  25. pa.ag
    @peakaceag
    25
    cleanup.pictures
    Remove any unwanted object, defect or even people from your picture in seconds
    Source: https://cleanup.pictures

    View Slide

  26. pa.ag
    @peakaceag
    26
    Ready for some scary stuff in-between?
    “We propose a new method based on a diffusion model (DM) to reconstruct images from
    human brain activity obtained via functional magnetic resonance imaging (fMRI)”
    Source: https://pa.ag/3ZBaHvG
    Reconstructing visual experiences
    from human brain activity offers
    a unique way to understand how
    the brain represents the world,
    and to interpret the connection
    between computer vision models
    and our visual system.

    View Slide

  27. pa.ag
    @peakaceag
    27
    AI audio & music tools
    Krisp AI: Removes background voices,
    noises and echo from all your calls
    Cleanvoice AI: Removes filler sounds,
    stuttering and mouth sounds from
    podcasts or audio recordings
    Podcastle AI: Studio-quality recording,
    editing, and seamless exporting –
    in a single web platform
    soundraw.io: Royalty-free music,
    AI generated based on mood,
    genre and length.
    Descript.com: To write, record,
    transcribe and edit video and
    podcasts
    Soundbite AI: Turn new and existing
    audio and video content into ready-to-
    edit social media posts, and summaries.

    View Slide

  28. pa.ag
    @peakaceag
    28
    Want to boost personalised interactions with customers?
    Maverick records once but automatically personalises for all, sending videos at scale.
    Source: https://pa.ag/3ZSIZdu

    View Slide

  29. pa.ag
    @peakaceag
    29
    MURF.ai as solution for challenges in video production
    Going from text to speech with a versatile AI voice generator
    ▪ MURF is an AI tool that offers realistic voices
    to use for voiceovers
    ▪ Depending on your contract subscription, you
    can choose between 20 languages and over
    120 voices
    ▪ Cost effective way without having to invest in
    equipment
    ▪ The tool is very easy to use and scales well
    Challenge:
    A client would like to use a
    video in different regions but
    does not have enough local
    employees to voice the scripts.

    View Slide

  30. pa.ag
    @peakaceag
    30
    The Peak Ace team used MURF.ai for a cost-effective
    localisation of video assets:
    Step 1:
    Transcribe the original script
    and localise it into the right
    language
    Step 3:
    Input your localised text
    and export
    Step 2:
    Decide on the voice you'd
    like to use
    Step 4:
    Add the new script to your
    video and adjust until it sounds
    perfect
    Original New French version

    View Slide

  31. pa.ag
    @peakaceag
    31
    Gen-1 by Runway in particular looks crazy cool:
    Just look at how it transfers the style of any image or prompt to every frame of a video:
    Source: https://pa.ag/3J6GDRA
    Source Video Generated Video
    Driving Image

    View Slide

  32. AI for:
    (More) Productivity

    View Slide

  33. pa.ag
    @peakaceag
    33
    Compose AI: A Chrome extension to make you fast(er)
    A free Chrome extension to automate your writing
    Source: https://pa.ag/41TToay

    View Slide

  34. pa.ag
    @peakaceag
    34
    ExcelFormulaBot: turn your problem into a formula
    Transform your text instructions into Excel formulas in seconds
    Source: https://pa.ag/3ythYlb

    View Slide

  35. pa.ag
    @peakaceag
    35
    Otter.ai does almost the same – and more
    Your meeting assistant that records audio, writes notes, automatically captures slides,
    and generates summaries:
    Source: https://pa.ag/3LhCcWC
    Word of advice (and caution):
    It's super practical, but
    getting consent to record, among
    other facts, is super important –
    especially since some services can
    automatically send meeting
    summaries by email afterwards!

    View Slide

  36. pa.ag
    @peakaceag
    36
    Need help to deliver a (visually supported) pitch?
    Try Tome app's approach to AI-based generative storytelling:
    Source: https://pa.ag/3FhIPEz

    View Slide

  37. ChatGPT… yup!

    View Slide

  38. pa.ag
    @peakaceag
    38
    Some useful Chrome extensions worth checking out
    Obviously, what is useful for you heavily depends on your goals and use-cases:
    Source: https://pa.ag/3Fc28iF
    My current favourites/recommendations:

    View Slide

  39. pa.ag
    @peakaceag
    39
    AIPRM for ChatGTP can make things easier
    Access a curated selection of ChatGPT prompts specifically designed for SEOs,
    marketers, sales, support, copywriting and more:
    Source: https://aiprm.com/

    View Slide

  40. pa.ag
    @peakaceag
    40
    Does ChatGPT remember what happened earlier in the
    conversation?
    Source: https://pa.ag/3FfiQgW
    While ChatGPT is able to remember what the user has said
    earlier in the conversation, there is a limit to how much
    information it can retain. […] up to approximately 3000
    words (or 4000 tokens) from the current conversation –
    any information beyond that is not stored. […]
    ChatGPT is not able to access past
    conversations to inform its responses.
    Always include all important parameters as
    well as style info, etc. in the original prompt.

    View Slide

  41. So, where are AI & LLMs
    headed in the future?

    View Slide

  42. pa.ag
    @peakaceag
    42
    Amazon entered the game with impressive first results
    Amazon has released a new language model that outperforms GPT-3.5 on ScienceQA
    by 16 percentage points, or 75.1% to 91.6% accuracy
    Source: https://pa.ag/3Fl5ZKe
    Amazon researchers came
    up with Multimodal-CoT,
    which combines visual
    features in a separate
    training framework, to
    reduce the effects of these
    mistakes [hallucinatory
    reasoning patterns]

    View Slide

  43. pa.ag
    @peakaceag
    43
    Language models using APIs for even better results
    This paper shows that LMs can teach themselves to use external tools via simple APIs
    Source: https://pa.ag/3T1Nq3z
    Toolformer: A model
    trained to decide which
    APIs to call, when to call
    them, what arguments to
    pass, and how to best
    incorporate the results
    into future token
    prediction.

    View Slide

  44. With one significant goal in mind:
    improving themselves
    Models will generate their
    own training data

    View Slide

  45. Because today’s LLMs make stuff up
    Models will fact-check
    themselves

    View Slide

  46. pa.ag
    @peakaceag
    46
    Planning for Artificial General Intelligence and beyond
    “As our systems get closer to AGI, we are becoming increasingly cautious with the
    creation and deployment of our models.”
    Source: https://pa.ag/3mxjSOU
    The first AGI will be just a point along the continuum
    of intelligence. We think it’s likely that progress will
    continue from there […] If this is true, the world
    could become extremely different from how it is
    today, and the risks could be extraordinary.
    A misaligned super intelligent AGI could cause
    grievous harm to the world; an autocratic regime
    with a decisive superintelligence lead could do
    that too.

    View Slide

  47. pa.ag
    @peakaceag
    47
    The best part of this OpenAI post though?
    Source: https://pa.ag/3mxjSOU
    Successfully transitioning to a world with
    superintelligence is perhaps the most important—
    and hopeful, and scary—project in human history.
    Success is far from guaranteed, and the stakes
    (boundless downside and boundless upside) will
    hopefully unite all of us.

    View Slide

  48. ...hopefully.

    View Slide

  49. Are we really ready for this?

    View Slide

  50. Not sure, but I certainly am ready
    for dinner! Let‘s do this! #food

    View Slide