It takes 3 FANG engineers to change a light bulb.
First one to write a design document.
Second one to align with key stakeholders.
Third one to deprioritize it from the roadmap because it’s a P3
@GergelyOrosz
I’ve had a hard time convincing some back end devs that front end is hard. But back end devs who do try and develop a complex FE application begin to appreciate the challenges.
@stanislavfort
@ylecun
@nisyron
I have a hypothesis for why these kind of prompts work. When you ask for a direct answer it might sample the wrong answer at the start and then spends the rest of the answer trying to justify its mistake. Any prompt that makes it delay the answer till the end will help.
@rakyll
A lot of blame has to go to perf evaluations and OKR processes that have long since stopped reflecting reality but don’t change. Just as how leetcode interviews are now gamed so heavily that the signal they provide keeps getting weaker. But it doesn’t seem to change how we hire.
@tszzl
Yeah but this can be misleading. Facebook can do this because it has always had world class monitoring , logging and fast roll backs. Those things are not optional.
@clairlemon
Correlations can surface interesting phenomena but acting on the basis of correlations without a causal model can lead to spectacular mistakes. We don’t even know what IQ really measures.
@sharifshameem
But the withdraw money button needs to be fixed too. I guess we can call this a new style of conversational programming. Where previous input and output are hints that help with a new corrected prompt. I wonder if GPT-n will know that money can’t be negative.
@Advaidism
So Madhya Pradesh with a mortality rate of 14 % is missing huge numbers of cases and so is Gujarat. Kerala is doing well on testing as indicated both by mortality rate and tests per million.
@GergelyOrosz
Front end also involves a lot of design and product thinking. It cannot all be done by the designer / PM. Making a smooth and delightful UX is actually hard. For eg Spotify vs the rest. Telegram vs the rest.
@abacaj
How does it work though ? How does a dumber model generate data that leads to a smarter one. Or does the delta come from filtering and curation that happens via the generation process.
@clairlemon
That doesn’t refute his point. He’s saying that correlation is not always as simple as it looks and the models are more complex in some cases. Nobody can disprove his claim because nobody really knows how intelligence works. But we do know how cigarette smoke causes cancer.
@antoniogm
@Jonathan_Blow
Twitter is not the right medium to have this discussion. The freeing of the individual has a historical context. Many of these obligations were enforced in oppressive ways by outdated institutions. So the question is how to foster duties in ways that don’t oppress people.
@pplsartofwar
@PAstynome
China converted population from a liability to an asset. And that’s the source of its achievements. And the quality keeps going up. Western countries have stopped investing in human capital and infrastructure. Plays out over decades.
@scarletinked
Teams and orgs at large scales regularly get bogged down in a morass of technical debt and fire fighting. The claim that code is not the hard part of engineering is simply not true.
An interesting piece of code that I won’t write again.
An app had poor App Store ratings. There was the bright idea of asking users for ratings only if the app did not crash recently.
A thread 🧵
@devdcdev
@abacaj
Scraping prices off competitors. I also know someone who built a price scraper in a hackathon at their company and got a promo. Then it became a whole team. One of the most senior ML ppl I know worked on this mainly.
@Advaidism
Good rule of thumb is that higher the mortality rate the more number of cases that go uncaught. Places with the most testing like Germany South Korea have mortality rates close to 1 %.
@GergelyOrosz
@mattrickard
I really miss the one off sessions. There’s only so much you can learn from documentation or reasoning. There are many things you pick up via osmosis. The lack of one off pairing in Big Tech actually holds them back.
@patio11
@paulg
To add some more nuance - Its possible that it doesn't move any of the KPIs everyone is focused on and any mistakes would lead to a big hit on key metrics.
@paulg
“Political opinions” is doing a lot of work here. It can be anything from Nazism to views on taxation. There are political opinions that pose existential threats to large groups of people. Whether those who want to enact them are good people or not isn’t relevant.
@patio11
Could have played it backwards and reused most of the work. Whenever something like this happens I’m sure there was someone in the room who felt the same way but couldn’t speak up
@Carnage4Life
Why are they assuming that 100 % of the money invested in an emerging technology will be returned. This is more of an arms race. If you don’t invest you get left behind. It’s about defending existing revenues.
@gonglei89
He’s basically saying that they built entire world class cities as Potemkin villages. It’s worrying that these are the people taken seriously in the establishment
@fchollet
I feel there’s a divide between those who want to develop an academic understanding of what intelligence is and those who want to make money out of it. Jobs can be atomized and repetitive. One doesnt necessarily need the full gamut of intelligence to automate a job.
@brajendra_rawat
@GabbbarSingh
@Rohan_29nov
Because manufacturing at scale is difficult. People scoff at China but making 1 of something is not the same as making 1 million. Requires an entire eco system of materials people and expertise.
@EMostaque
We need to distill it more so it can run on consumer GPU. I remember deep mind had some paper on RETRO or something that allows the model to be split into a cacheable database component and a smaller model.
@revhowardarson
IPCC projections make a lot of assumptions about emissions reductions and sequestration. Plus we don’t fully understand if / when feedback loops kick in.
@tmas42
@GergelyOrosz
I already decided to take the big hit in comp and move to the UK instead. And this was before the current mess. Indians need to carefully weigh the trade offs of moving to the US.
@clairlemon
And taleb knows about this from experience. Many traders blow up when they use models that exploit correlations that they think will always hold up until they don’t because they don’t know why it exists
@rakyll
ZeroDha is a great example of a company that has avoided all of these pitfalls. Is a billion dollar company with a very small team of engineers.
@pplsartofwar
@angelusm0rt1s
There’s something to American culture where being too level or knowing too many details is considered low status. Probably explains a lot of the decline of companies like Boeing
Rolling out features on mobile apps can be tricky.
Here’s one of the worst bugs I’ve ever shipped to production.
When working on hotel bookings we decided to parse a description string on the client to show if certain amenities were available.
To ship dozens of user-facing features every week to hundreds of millions of users, while they are all at different stages of rolling updates, is both art and science. Also you need to A/B certain releases.
Find out how we do this all
@JioCinema
@abacaj
Lets think of it as curve fitting. So I have fit a some curve to my data. Now I sample some more data points from the curve I just fit. Now I refit the curve.
This seems problematic ?
@abacaj
It’s important not to be 100 % swamped with work. Only take enough work to occupy you 70 % of the time. In the remaining time quickly ship stuff like this that moves the needle. That’s how people get ahead. Not by grinding on pre decided items.
@atroyn
"In December 2018, declassified documents revealed that the CIA made six dogs run, turn, and stop via remote control and brain implants as part of MKUltra."