jump to navigation

Blast to the Past – Past Drives Future Growth (Notes from June 17 to June 23, 2019) July 9, 2019

Posted by Anthony in Uncategorized.
trackback

I’ve included a few repeat guests but from different podcasts. Fun to see how different the discussion can be with a different host. So, if you’ve been following along, hope to hear that you recognize Eric Topol and Joseph Jaffe’s names.

I was quite intrigued by Archit Bhargava’s work at Niantic in marketing their games further, along with the creative process behind how they started. Niantic did Pokemon Go – and what went in that from starting with the maps and what game may have worked (some… ~10 years later). The Crew communication app also had a fascinating introduction story to get to where they were – from sitting in a tequila bar coming up with the name to finally developing an enterprise communications app.

Then, there was a number of data science-centered episodes (of course). A16z had a discussion in ML and AI for medicine – how we see it, where it stands, where it’s weak and should improve. Back to the future was also a method for the Endangered Language Project where 2 contributors were on Data Skeptic discussing their research while at USC going through NLP on languages that are losing speakers/writers/people to pass it on.

One of the most exciting and energetic guests were the co-founders of Bulletin Space, though. Two women who eventually decided to make their brand a woman-centric platform focused on products by females for females, and providing them space to do great work.

Hope everyone enjoys!

  • Archit Bhargava, Head of Global P/Ming at Niantic Inc (Work of Tomorrow)
    crw0zgjvaaaig6k

    • Discussion of Harry Potter game after Pokemon Go
      • Social aspect of exploration
      • Partnering with cities and maps points of interests
    • Revenue model – in-app purchases vs sponsored placement of gems / pokemon
      • In Japan, partnered with McDonald’s
      • US – Starbucks and gyms
  • Joseph Jaffe (@jaffejuice), author of Built to Suck (Wharton XM)
  • AI and Your Doctor, Today and Future (a16z 6/13/19)
    deep-medicine-cover

    • With Eric Topol (@erictopol) (author of Deep Medicine) – cardiologist and chair of innovative med at Scripps Research, Vijay Pande (@vijaypande)
    • Didn’t expect AI and medicine to be such back to the future – outsourcing so many things that could get us back to the 70s and before
      • Doctors spent more time with humans, patients due to big business, EHR, admin wants more “efficiency”
      • On Twitter – kid drew a drawing of going to doctor – it was a doctor with his back turned typing on his computer
    • More tech is less computer – fundamental problem, not even drawing eye contact
      • NLP can already liberate time, some UK emergency rooms, as well – eliminate the distraction and data clerking
      • Encouraging to have the speed/accuracy for transcriptions, ontology and organized data
    • Google AI  on improving the voice processing
    • Min discussion / comm b/w doctors for treatment/diagnoses
      • In his book, he went through his knee replacement surgery – orthopedist wasn’t in touch with his congenital condition
        • Logistics and coordination for computers – for thyroid cancer, maybe would need endocrinologist with oncologist
      • How do computers know things that we don’t know now? Complementary – big data has appetite for it (humans – contextual)
    • Radiologists have a false negative rate of 32% – ground truths for x-rays / scans that it won’t miss – basis for litigation (missing some)
      • Best use of time for doctors would be understanding and discussion with patients
    • Diagnosis in general – once trained, doctors are wedged into their diagnostic performance for their career
      • Kahneman’s System 1 + System 2: if doctor doesn’t think of the diagnosis in first 5 minutes, they have an error rate of 70%+
      • ML is reflecting system 2 since it’s trained off of doctors doing system 2 – but with an aggregation of 1000s
    • Up to 12mil errors in medicine a year – can improve upon this, easily
    • Negative components potentially:
      • Can’t trust unilaterally, need oversight
      • If FDA approved, have to watch for cohorts – deep liabilities, ethics, privacy issues – have to be tradeoffs and considered
    • Rolling out AI – NHS is the leader in genomics (ER rooms without keyboards), China with scale and advantages – data on each person
      • Limitation of data in the US and otherwise, no strategy here as well – other countries have a lot of resources spent / proposed
      • Education and training has a full wing for AI in the NHS
    • Professional organizations have not been forward thinking – maintaining the reimbursement for their members
      • Worst outlier, outcomes and mortality – worst, especially due to spending ($11k per capita)
      • Naivete of diet and having the same diet – not just glucose responses, but triglyceride tracking – non-normative responses to same thing
    • Wiseman Institute in Israel cracked code on promoting health – glucose, lipids in blood – eventually see outcomes / prevention by diet
      • Numbers of level and data for each individual – specific bacteria and the sequences, physical activity, sensors for stress, sleep
      • Need hundreds of Ks of people to learn more and on a broader spectrum
      • Can give retina picture to Intl retina expert and it’s 50-50, but an algorithm is 97%
        • Polips in colonoscopy, K level in blood thru smart watch w/o blood
    • Why did people go into the medical profession in the first place?
      • Care, helping people and seeing people – but now have highest suicide, burn out and everything
        • More depressed which leads to more errors and cycles
    • Could do remote monitoring, for instance, for all of non-ICU patients
      • Hospitals won’t allow it, they’d be gutted. Hospitals are problematic – 1 in 4 get injured or sick.
      • Only quick adoptions are enhancement of revenue (think: robotic surgery rooms)
      • Comfort of own home – can sleep, see loved ones, clearly cheaper (average is $5k / night)
    • Cardiologists thought you’d have to look at the QT interval – only 1 part
      • Flunked with the algorithm – Mayo Clinic wanted all data and use the full cardiogram
      • AI / ML already have a great driver detector
    • Easier for machines to do new things with no regulation – rare cell detections, genomes
      • Imagination is our limit / machine limit (unsupervised learning limited by annotation, for instance)
      • Prediction has not done as well as classification, clustering (uses his stepdad as an example, who was resurrected)
    • Not there yet for multimodal algorithms – doctor doesn’t have to do typing – AI figures out diagnosis
      • When you go to see a doctor, you want to be touched – the ‘experience’
      • He doesn’t use a stethoscope anymore, he uses a smartphone ultrasound for EKG – shows the patient in real-time
        • Tools of the exam may change, but interaction will be intimate still – have to get back to this
  • The Death of a Language, Endangered Language Project (Data Skeptic 6/1/19)
    maxresdefault

    • USC students research Zane and Leena, CS / App Math and CS / Cog Sci
    • Using unsupervised learning to assist classifying pho (basic unit of sound in speech) names of endangered languages with PyTorch
    • Last living speaker of a language dies – globalization has rid the world of a number of languages (Latin, inc)
    • Helping linguists and a sociological effort to carry on the language – Zane’s father speaks a dying Italian dialect (Venetian, so not living)
      • Very similar to Italian just from listening to audio but has different conjugation / words for some
    • 3.5 hours of audio from 4 speakers collected by their professor in an area near the Northern Italy
      • Most valuable for research – more speakers to improve dataset
      • Output as recommended start and stop times, unsupervised labels for the times – rec time for pho name
    • Slicing audio file into many small segments, labeling them and then combining the adjacent segments
    • Built classifier with an NN – series of vectors (condensed, auto-encoding of audio data)
      • 7000 spoken languages but estimated that half will go extinct by the end of the century
      • Manual transcription is tedious, so hoping it will assist linguists in proscription
  • Three-Legged Stool, Chuck Akre – Akre Capital founder (Invest Like the Best, 6/18/19)
    chuck-akre-1024x1024

    • Managing $10bn after forming firm in 1989
      Being in “one-stop town” quality of life, not being disturbed by outside events or people

      • Distracted or curious by friends, their thoughts – he spends a lot of time reading vs screens
      • Thing that disappointed his son, a tenured professor without luxury of being highly select
        • best students (those that got A’s) wanted to only know what would get them an A, rather than have curiosity
    • 3-legged stool as more stable than 4 legged – can deal with uneven surfaces
      • Rates of return in common stocks were higher than any other asset class in unlevered case
      • “Once a guy sticks his hand in your pocket, he’ll do it again” – human behavior happens to be antithetical sometimes
      • Legs: quality of business enterprise, quality and integrity of people running it, what is their record for reinvestment and opportunity for it
    • Own exceptional businesses, don’t sell them
    • Bandag – “tire business” but looked at returns on capital and they were 3-4x everyone else
      • Wanted to figure out what business it was in – Marty Carver (founded by his father) was a different feel
        • Retreading bus and truck tires, early 70s had tanked oil + petroleum so dealer’s had skyrocketed costs
      • Bandag passed on savings to their dealers in the evidence that they had to reinvest the money in their stores – new ones, franchises
      • Created huge dealer loyalty – trucks/buses’ tires are constructed so that they retread 2-3 times
    • Weren’t smart or quick enough to get into FANGs or otherwise – couldn’t figure out the value creation in quick span
    • Mastercard March 2010 – during Dodd-Frank, Congress act, Durbin amendment – Mastercard/VISA selling at 10-11x
      • Discovered operating margin on returns over capital “not enough words that were superlative enough” – still above average after cutting 50% twice
      • Everybody wants some of that – jamming every expense into the income statement to reduce how good margin is
        • What’s causing that? How do they have it?
    • Wall Street – wants transactions in general (his biz model – compound capital) – create false expectations on earnings estimates
      • “Miss by a penny, beat by a penny” – gives them opportunities since markets behave irrationally
      • Dollar Tree as 3 major players and that was it
    • How do you measure whether you’ve been a success at running this business? Or we hit our earnings target or board goals, etc.
      • Impact of compounding economic value per share. Not trained to do that.
      • O’Reilly as duopoly and buying and deleveraging – capital allocation change / International Speedway
    • Growth of antennae – American Tower Corp (buying in Africa recently) – 5G won’t be here soon, but they’re acting as gatekeeper
      • Microsoft as the OS toll
    • Collecting datapoints and making judgments on them in general, whether you’re an English major, pre-med, investment management
      • Reading business biography learning behaviors
    • Land conservation plan / donations
    • What is the source of pricing power for each company?
  • Named Entity Recognition (Data Skeptic, 6/8/19)
    • Entities as core features of a sentence, idea
    • Text file analyzing or software doing a good job of named entity
      • If you give labelings, some are from computer vs English majors (Turing test)
        • Using SpaCy for NER – hard problem, different expectations but not great – just good
    • Chatbots seeking NER – flight example, for instance – pull out things that are mentioned
      • City is destination, airline mentioned
    • BERT can do NER pretty well – Google Assistance and chat interfaces have been improving
      • Semantic web projects can pull entities out of documents and connect them in knowledge graphs
      • Transfer learning – pretrained model on generic model and use that as jumping off point
  • Carmax: Way Data-Science-powered Car Buying Should Be (BD Beard, 5/28/19)

    • Tod Dube, Chief Architect for Data Science at Carmax
    • Adding 1 store per month, no. 3 wholesaler (to Barrett Jackson, eg), $18bn in rev, #1 used car
    • Determining pricing is through ML, but now omnichannel – looking/exploring cars interactions
      • Customer service buying exposure
    • Changing how data scientists go about their job – laptops with minimal compute power, governance issues trying to fit onto laptop
      • Analytical leadership to push tech to do things better – how to make availability
        • Security, data losing, privacy, model
      • Shift data and move data around but if data moved in 3 weeks – how can you iterate easily
    • Architectural changes from laptop / personal side to service and data warehouse pulls / data centers
      • Azure service response – pick use case, can’t swallow the elephant (replatform rec that were done today – handwritten code)
      • 2 week sprints for changes before – different cars, reasons, prices and availability
        • What tools could help? SaaS, subscription to bridge the gap – Had python (Jupyter, Spyder, Anaconda) / R
      • Started a data lake because of use case
        • Had to pivot and find data scientists (Type A – analytical, business; Type B – data engineer, why model is necessary for data)
    • Consulting partners as unsung heroes to figure out how to build out a team or look at problems
    • Spark as a Service, Spark as data lake, DataBricks (Delta Lake), Azure customer
      • Will auto finance almost any cars, call centers – better enabling customers in financial choice
    • Walk-on song for conference: I’m Not Afraid by Eminem
    • Spends money on tech, iPad Pro new now, MacPowerUsers (how their workflow is)
      • AI, weather on sprinklers for rain predictions
  • Ali Kriegsman, Alana Branston, co-founders of Bulletin (Wharton XM)
    bulletin

    • Switching from platform with competitors like Etsy, Bazaar to drive sales, initially
      • Had creative, original content and hooked brands up with some unused channels/media
      • Asked brands/customers after > 100 brands – ‘peak’ initially
      • They talked about how valuable, but expensive, physical space was – pop-ups individually, Brooklyn Flea, etc
    • Decided to do pop-ups in big parking lots – ineffective, felt the heat – literally (12k sq ft in parking lot outside)
      • Shrunk it and rented out a front space from a bar – charged $300 * 30 brands for pop-up for a weekend
        • Worked for both brands and them – realized they could do that ~10 months
      • Grinding for the year, made money equivalent to month of prior sales, but 7 days / wk wasn’t scalable
    • Finetuned branding for pop-ups to female founders, female products (had men originally)
      • Best products at time were stamped necklaces, ready-to-wear clothing increasing
  • Sucharita Kodali (@smulpuru), Retail Analyst – Forrester (Marketing Matters)
    forrester
  • Danny Leffel, CEO, cofounder of Crew (Wharton XM)
    0ed3222e-7f0e-4606-9e27-789e09b4f110-1548112861887

    • Communications app for everyone on professional page
  • Bryan Murphy (@bryanpmurphy), CEO of Breather (Wharton XM)
    breather_website_logo

    • Talking about the optionality to get working space

 

 

 

 

  • Dan Widmaier, CEO cofounder at Bolt Threads Biotech (Wharton XM)
    nav.logo_

    • Using spider silk and attempting to synthesize stronger proteins for apparel, clothing
    • Tie was the first – needed a quick demonstration
    • Have gone on to other materials, solving environmental waste of apparel
Advertisements

Comments»

No comments yet — be the first.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: