sunpazed

Rolex are the De Beers of the mechanical watch world. Scarcity and exclusivity is a great way to market luxury. A Rolex Submariner in the 80s was about $1,000 — these days it’s closer to $12,000. A new Daytona is twice that. While they were never cheap, they were affordable for a working professional.

I’ve been in product management for over 20 years, and I agree with most of what is said here. Ratios matter. CPOs end up empire building with an army of PMs. The organisation ends up with PMs owning a slice of the customer experience such as a single page, optimising CTRs on buttons, with no meaningful contribution to the product strategy or vision. It’s hard to be a Product Manager without owning a product.

Try a quantized model on llama.cpp — I have an M1 Max 32Gb and it’s running fairly fast in 20Gb of memory.

See this related project: https://github.com/rvdveen/epaper-slow-generative-art

Firstly, ethical reasons. Secondly, the act of experimenting in a model organism is a perfectly acceptable way of generating science. Rats have helped humans develop cancer drugs, HIV antiretrovirals, and even the yearly flu vaccine.

Yes, LUTs are your friend. I used to code scene demos on the C64 back in the 90s.

I’m writing a pinball game in assembly for the original Gameboy, and recently had to hand-craft multiply, division and cosine routines to trade off accuracy for spare cycles. Something like this would have come in handy!!

27 million people live on a continent the size of North America!

Australia’s geography and climate plays into the housing crisis. We have concentrated our population density in only a dozen or so locations (ie; major cities like Melbourne, Sydney, Brisbane) as Australia is only habitable at scale along the coast.

Australia has also doubled in population size over the last 40 years without corresponding investment, putting significant pressure on the housing market.

OpenAI less so, small community trained open source LLMs more so. We’re already seeing this on HF with community fine-tunes on Mistral and Llama2.

In the same way search transformed knowledge augmentation, LLMs will transform skills augmentation.

Forget about the things you know to do well, instead focus on all the new skills LLMs will unlock for you.

This reflects my anecdotal findings in a very narrow domain: GPT-4 consistently resolves complex SQL from natural language, whereas GPT-4-Turbo is hit-and-miss, and similar to 3.5-Turbo in performance.

Here in Melbourne we experienced six lockdowns, totalling 262 days, with 111 days being the longest consecutive stretch. At times we couldn’t travel for more than 5km from home, and only for certain government sanctioned reasons.

https://www.abc.net.au/news/2021-10-03/melbourne-longest-loc...

Initially I agreed with the approach given the fragility of our hospital system, the escalation in cases, and the lack of a vaccine. However, this was not without personal cost — once the longest lockdown was over, it was clear that most people psychologically struggled to transition back to pre-lockdown norms. My youngest child had the hardest time of all of us physically re-engaging with others in society.

Was it worth it? Perhaps. Did it cost us? Absolutely.

All the individual ingredients were around for years prior to the iPhone — we had apps, music and even video on smartphones.

Apples magic was to greatly simplify the user experience, and make smartphones and their content more accessible to the masses.

Well, Metal can only allocate a smaller portion of “VRAM” to the GPU — about 70% or so, see; https://developer.apple.com/videos/play/tech-talks/10580

If you want to run larger models, then CPU inference is your only choice.

I’m in my mid 40s, and honestly I have more ideas than I ever had in my 20s!

However, the big difference, is energy and motivation.

I was a machine in my 20s, but these days a late night hacking session ends with me paying the price the following day. Execution is everything.

I’m in my late fourties’ and have witnessed only a handful of transformative technologies in my time. Nothing in the last 15 years has given me that “tingly feeling” of excitement — you know, that feeling that we’re on the cusp of something transformative — than the recent progress in AI.

While the new AI frontier might be led by prohibitively expensive (and closed) large language models, we’re also seeing great grass-roots progress at a smaller scale with modest models trained by the developer community. I trained a baby GPT the other day using llama2.c for my own use cases.

It’s Linux vs Sun/Unix all over again.

Here are my kids playing my childhood 2600 from 1982 https://www.instagram.com/p/CIJ0aeTJAuo

"The medium is the message" should be the take-away here, ergo; not what Generative AI does, but how it shapes us. The Internet augmented our connectivity to one-another. Search augmented our knowledge. Generative AI is augmenting our skills.

I can get this running on my M1 Mac Pro (CPU mode) with a few tweaks in the cv2win32.py file. I'll submit a change request.

I actually quite like it! It’s worth a purchase

In my last role, I built an education reviews site at scale — challenge is remaining objective given that your audience are actively airing the dirty laundry of your paying customers. Commercially it didn’t stack up, so we had to monetise the traffic through display ads.

Same case with Glassdoor — dissing paying customers through your audience is never good for business.

Mobile site scrolling horizontally within the viewport, ie; scrolling is broken for mobile break-points.

Perhaps they’re the outlier, but Apple customer service is great. I’m only ever a few minutes away from talking to a sales or support person.

I work for a call centre observability startup (operata.com) and my take is — call centres aren’t going away anytime soon.

In fact Zoom, Microsoft, and Salesforce have all recently announced they are entering the Call Center as a Service space (AWS already have their own call centre PaaS).

In the consumer space we’re seeing further consolidation and homogenisation of products/services. It’s now so easy for me to churn my energy retailer, my insurance broker, my phone provider, etc.

Outside of price, the only other competitive differentiator will in-part be the customer experience.

Websites, chat and BOTs are great for standard workflows. For reasonably complex customer queries (either sales or support) nothing yet beats voice or video with a human.

I can’t help but recall this Cheyenne Dialysis satire whenever I watch Zuck lead a keynote; https://youtu.be/E4jQe9h5v5E

Adobe charged me an “early termination” fee of $43.99 after 6 years of continuous subscription! https://twitter.com/sunpazed/status/1561679328995250177

Support didn’t help, so I had to vent on Twitter before they refunded the amount.

Not happy with their obscure and deliberately misleading subscription model.

Search, data normalisation, equivalencies, recommendation engines, supply chain analytics and reporting, etc.

Having a well defined ontology helps power lots of data products, and even build new ones.

Our family camera was a Kodak Disc Camera. The crappy grainy quality was offset by the speed and convenience of swapping discs. Lots of family moments would have been forgotten, so the trade-off was worth it.

Only problem was price (more expensive to buy and process), and availability (hard to find disc film in a country town).

Ah cool, I’ve replaced so many of these in the past!! If I recall, RetroKit and a few other stores sell more reliable modern PSUs.

Both the C64 and C64C were released in Australia. Here’s my original C64C from the mid 80s for reference: https://imgur.com/a/5J9lSgK

Still works perfectly.