Appreciated the response! I noticed the same when I ran tau2 myself on gpt-5 and 4.1, where gpt-5 is really good at looking at tool results and interleaving those with thinking, while 4.1/o3 struggles to decide the proper next tool to use even with thinking. To some extent, gpt-5 is too good at figuring out the right tool to use in one go. Amazing progress.
Correct, and with reasoning, the ratio is totally off. As others have pointed out, actual usage is way higher (much more than 3-5x) than the estimation in the article, which is probably for very trivial users.
The estimation for output token is too low since one reasoning-enabled response can burn through thousands of output tokens. Also low for input tokens since in actual use there're many context (memory, agents.md, rules, etc) included nowadays.
While there're interesting findings here, https://arxiv.org/pdf/2502.03373 (also with a lot of good findings) suggested some contradicting theory on the critical mass of training process/data for the sake of reasoning capability.
Strangely, deepseek has been always a prominent name in open source LLM community since last year, with their repos and papers - https://github.com/deepseek-ai. Nothing of it is really quiet except that they probably burn 1% of marketing money compared to other China LLM players.
There’s an interesting dynamic here. What value of OpenAI to use for conversion calculation? 86b round is pretty much dead if they move to MS, yet 29b is too low (even 86b is low in terms of future potential). And what kind of upward room there will be?
I have no doubt that MS can spend billions in cash or RSU to compensate all of them, but I do believe there’s some gotcha if exodus actually happens and MS might not be so generous in throwing millions of dollar cash for a general backend engineer recently joining OpenAI.
I'd recommend Julie Zhuo's The Making of a Manager. She's got a twitter https://twitter.com/joulee where you could check out her sharing of managerial wisdom to have your own gauge.
This book focuses on the practical side like sharing useful feedback, smart recruiting strategy and meeting optimization, all towards the goal of greater outcome and amplifying team success, instead of just more activities entailed by conventional manager model.
On personal anecdotes, I've used the same internet identity across all platforms and it's not hard to locate my real life identity from there...since like 20 years ago when I got on the internet. So all the good reputation and bad reputation followed me through, and surprised me in best and worst way.
There're times when I wished I had concealed myself saying certain things (or just haven't said it), and times when I felt the joy of recognition when people connected dots about me and became closer friend. Now as I'm getting old, it's just a habit, no longer about risks or rewards.
Yet, both younger me and older me will probably be very indecisive about doing this if they've never done it and are told about all the fortune/misfortune this has brought to me, including meeting my wife and thinking about suicides.
"A common question I get is about work/life balance. And I view it a little different too now. At Facebook I could get by barely working a few hours a day. But the job was so unsatisfying that it spilled the frustration in the “life” part"
This is the golden piece for me. I too recently left a big corp to a smaller place, expecting to sacrifice WLB and my free time - yet I ended up feeling happier and spending less time consuming short-form online contents, to the point that I even have more free time now! It never occurred to me that my self-indulging online addiction is actually a coping mechanism towards the unsatisfying and meaningless daily job.
Yet you might be surprised to know that, in China, all those "offline games" are categorized as online games because they all can be distributed online, when it comes to import reviews. No game is safe from this, really.
With Google doing front loading(33/33/22/11) and other companies also shifting away from conventional 25% per year vesting to make first few years more attractive, they gotta catch up to stay competitive. Uber recently did the same of removing cliff.
I'd like to provide an personal take on this - this will be a bad deal until Uber can be a super app doing 100x more things than it's doing today. So maybe never.
Why? Because exactly in China there was consolidation of food delivery apps into like 3 or 4 of them from 100+. But all of them are still burning lots of money from investors, except one - Meituan-Dianpin which built on this super app (first think of it as uber+yelp+groupon+tripadvisor+more) handling all kinds of life needs from food to labor service. Yet this super app barely started making profits recently after grabbing so much aspects of daily life, while its food delivery unit is still burning money, though contributing to the profiting units like ads.
The title misses the other, and more impactful/unprecedented change of giving everyone "exceeds" rating and skipping perf review for them (except for those high performers who will get separate review).
That does far more than $1000 to relieve everyone and make it fair for those who have to take care of others in the family during such time.
I’ve also read this book a couple times but sadly as a engineer in large and mature company I hardly find any opportunities to actually apply or strengthen the learning, then as time goes my learning just decays.
How are you tackling the issue? It’s not like that one can find distributed system projects easily.
This is indeed very interesting to see and made me want to share my stories. I've also scanned over thousands of books (mostly Japanese manga, in a way to digitalize them and make them persist) myself and ever wondered where tech could bring this digitalization to. I've spent more than 10,000 hours on this for more than a decade and will probably do it until I couldn't.
Due to the requirement for maximum restoration of manga contents, especially for those cross-spine scenes, the high speed camera-based solution like above or other places which focus on formal prints won't work quite well and that leads to two major ways in the trade for decades since some Japanese started this self-scanning thing around 80s:
1) Destructively break the book into unglued pages (which also invovle cutting the glue with heavy-duty paper cutter) and send them to scanner with auto feeder, (then try to glue it back if keeping the book is vital, at best effort).
2) Manually press each page on flatbed scanner (thus A3 scanner works better as it allows vertically scanning spine area usually with the greatest shadow) and scan each page using predefined rough area which would therefore include non-content area as shown in : https://imgur.com/a/2ITBlJg (left/right edge and spine).
Solution 1) works great in terms of efficiency but could be a pain for those who love to keep book still in the collection (though malformed anyway, and I myself might consider this when I get old and couldn't bear the effort). 2) would cost a lot more time (on average, #1 takes less than 30 mins per book and #2 takes at least an hour) especially for manually removing noises as shown above. I've tried some basic CV tools/scripts to auto cut at least the edges (middle spine could be troublesome and I'm fine with manually working on that), which work really poorly since content itself could also be largely variant if we try to determine "edge" by checking pixel distribution.
Hopefully before 2nd decade finishes, I either give in to destruction or find a perfect way to automate the noise-clearing process.
I've lived in different cities in China and US. 4-6 RMB to 1 USD in major cities like Beijing/Shanghai, and 1-3 RMB to 1 USD in less developed cities or sururban area. Less developed a city is, less RMB to USD in purchasing power for most domestic goods.
I did ask my current manager about this decision and he did mention that he wants to find a engineering project manager (we have PMs) who could technically manage projects with him so he's not overloaded. This person might become my manager in the future but not immediately (maybe after 1 or 2 years when he's going to assemble his own team).