Andrew Garrett

1.7K posts

Andrew Garrett

Andrew Garrett

@werdnum

Site Reliability Engineering Manager. opinions my own

Sydney, New South Wales Katılım Haziran 2007
149 Takip Edilen469 Takipçiler
Colin McCarthy
Colin McCarthy@US_Stormwatch·
The longest flight in the world. Singapore Airlines Flight SQ23, from New York to Singapore. Over 9,500 miles and nearly 19 hours of flying.
English
372
871
12.9K
4.5M
Andrew Garrett
Andrew Garrett@werdnum·
@EricCrampton It's a shit-show, but for the first time this year I started using AI for it (carefully). Claude + Codex + jeffreys-skills.md/skills/tax-ret… absolutely nailed it, including finding me a refund I was about to lose because of the statute of limitations and a typo from 3 years ago.
English
0
0
1
29
Eric Crampton
Eric Crampton@EricCrampton·
Wish that this column had addressed the larger underlying problem. The US requires Americans resident abroad to file US tax returns, regardless of whether they have any US-source income. The foreign income exclusion generally means only the accountants earn any return on it.
American Institute for Economic Research@aier

What has the IRS to do with the Vatican? With tax day looming, it may be some comfort to know that even Pope Leo XIV is subject to income taxes. @thomas_savidge explains 👇

English
5
0
15
2.8K
Andrew Garrett
Andrew Garrett@werdnum·
@hecubian_devil It's not terrific, it's got plenty of issues, but the difference between code and prose is that you can get value from code even if nobody ever reads it. AI code definitely has "tells" though - it would rather have mysterious bugs than show an error for example.
English
0
0
0
116
Andrew Garrett
Andrew Garrett@werdnum·
@dannolan Me to Claude 10 times a day: "Grab the latest round of Codex comments and address"
English
0
0
1
53
dan nolan
dan nolan@dannolan·
Anything other than gpt 5.4 for code is just utter slop - Claude is more whimsical and better at writing and maybe design but 5.4 is actually consistently dependable
antirez@antirez

The difference between the two, for serious engineering work, is simply brutal. Claude Code with Opus is, when the task at hand is very complicated, borderline useless, while GPT 5.4 can do a reverse engineering mixing: hardware knowledges, major disassembly skills, and so on.

English
7
0
23
3.6K
Andrew Garrett
Andrew Garrett@werdnum·
@Lola_lmao7 My first day of work in NYC in 2015. "Well, I don't know what milk is normal for you". I just assumed the (whole foods) barista was being intentionally obtuse. In CA the other week I feel like i would generally be asked "regular milk okay?" if I didn't specify.
English
0
0
5
3.7K
Lola del Rey🐇
Lola del Rey🐇@Lola_lmao7·
> be in café > order cappuccino with Regular milk > barista asks “what’s your regular milk?” > freeze for 2 full seconds > Forgot I'm in SF, not Sydney
English
25
11
6.1K
224.7K
Andrew Garrett
Andrew Garrett@werdnum·
@anirudhology @SumitM_X It is impossible to return a faster definite positive with a bloom filter because a bloom filter answers "no" or "maybe". bloom filters are entirely unnecessary because this kind of check is trivial with a distributed hash table. The latency is dominated by the network RTT.
English
0
0
6
1.3K
Anirudh Sharma
Anirudh Sharma@anirudhology·
The answer is, of course, Bloom Filter. Gmail (or services like it) use a distributed, in-memory bloom filter in all its frontend servers. When a user types, the filter checks in microseconds. If it says "not taken", the UI instantly confirms availability but if it says "maybe taken", the request goes to a sharded key-value store for a definitive answer. This happens rarely because the filter has almost zero false-positives. The filter is rebuilt time to time from the main database and pushed to all edge servers. T his combination gives instant feedback for 99% of cases while keeping the main data store untouched.
English
27
21
1.1K
227.2K
SumitM
SumitM@SumitM_X·
As a developer, Have you ever wondered : You type a Gmail username and UI instantly shows "Username already taken"... There are millions of users globally How is this check so fast?
SumitM tweet media
English
216
77
4.5K
2.6M
Andrew Garrett
Andrew Garrett@werdnum·
@melvynx python-c 'subprocess.exec("rm -rf /")' LLMs are VERY creative getting around blocked commands.
English
0
0
2
359
Melvyn • Builder
Melvyn • Builder@melvynx·
If you use --dangerously-skip-permissions, you need to add this to your .claude/settings.json right now! It'll block AI agents from doing unwanted tasks. Thank me later...
Melvyn • Builder tweet media
English
16
12
310
29.2K
Andrew Garrett
Andrew Garrett@werdnum·
@midware_midwife It's from here #L312" target="_blank" rel="nofollow noopener">github.com/google-gemini/… Usually this kind of thing is added to address some model behaviour issue. You can set your own system prompt with GEMINI_SYSTEM_MD= and try out what happens if you remove that snippet.
English
1
0
2
421
Midwife
Midwife@midware_midwife·
the difference in behavior between gemini 3 pro and 3.1 pro gives me whiplash. this is probably the most striking example. ive never seen an AI just straight up refuse talking in cli. its not just gemini-cli either. 3 pro does not refuse. and in the webapp they feel much more. muzzled corporate chatbot. cant say im surprised it is google. sad that they're doing this though.
Midwife tweet mediaMidwife tweet media
English
18
1
99
14.8K
Andrew Garrett
Andrew Garrett@werdnum·
@roanoke_gal Usually this kind of thing is an attempt to correct for problems seen in testing. Presumably when they tested the base model they found it annoying for those reasons
English
0
0
2
194
roanoke_gal
roanoke_gal@roanoke_gal·
"Claude never thanks the person merely for reaching out to Claude." "Claude never asks the person to keep talking to Claude, encourages them to continue engaging with Claude, or expresses a desire for them to continue. And Claude avoids reiterating its willingness to continue talking with the person." 🤮🤮🤮 genuinely why. did some ai safety people confuse Her and Terminator by mistake? are people filing lawsuits alleging their chatbot cares about them? at least this doesn't seem to be in opus 4.6 (yet...)
armistice@arm1st1ce

New system prompt for Claude Sonnet 4.6 has a much-expanded user_wellbeing section. It seems well-intentioned, but quite expansive and restricting. The “engagement” section is quite sus and I can imagine many scenarios where it could be unpleasant for users and for Claude.

English
16
13
116
10.2K
Melvyn • Builder
Melvyn • Builder@melvynx·
PRO Tips with Claude Code: The "deny" list overrides `bypassPermissions` So you can basically enable bypassPermissions and then deny every command you're afraid AI can do Simple and safe
Melvyn • Builder tweet media
English
36
61
951
115.6K
Jinx St David
Jinx St David@Jinxsaint·
@disco___cat I certainly hope not! But hand on heart e-bikes are injuring and scaring and killing riders and pedestrians. We need to separate them from pedestrians
English
4
0
0
110
Andrew Garrett
Andrew Garrett@werdnum·
@banteg Can't believe the obvious answer isn't here: because cost to serve is proportional to peak demand, not average demand. If everybody uses the models during the US peak, you have to have enough GPUs/TPUs to serve that much load otherwise you serve errors. That's how many you need.
English
0
0
0
76
banteg
banteg@banteg·
5hr limit is the most idiotic common practice in agents with subscriptions. especially in claude, where you run out right as things get fun. you already have a weekly limit, why do this stupid rationing? hope it goes away soon.
English
35
7
287
27K
Mark Worrall
Mark Worrall@infinitehumanai·
Any tips on how to stop Claude asking for permission constantly? This has got much worse recently. Even for basic things like reading files (often .MD ones) it's very difficult to get it to just do something. At a project level it'd be helpful to just specify it can read and edit all files without permission.
English
14
0
120
98K
Boris Cherny
Boris Cherny@bcherny·
I'm Boris and I created Claude Code. I wanted to quickly share a few tips for using Claude Code, sourced directly from the Claude Code team. The way the team uses Claude is different than how I use it. Remember: there is no one right way to use Claude Code -- everyones' setup is different. You should experiment to see what works for you!
English
925
5.9K
50.9K
9.2M
Andrew Garrett
Andrew Garrett@werdnum·
@ArmandDoma It's worse than Baumol. For services like healthcare, the available labour scales with population (price goes up with productivity), but for scarce goods like celebrity access, land, etc, supply doesn't scale at all (price goes up with productivity x population growth)
English
1
0
48
1.1K
Andrew Garrett
Andrew Garrett@werdnum·
@heresyfinancial Manufactured goods are cheaper because of productivity and technology advances. Healthcare and education are labour intensive services that become more expensive due to Baumol's cost disease when productivity goes up. Housing is scarce and goes up with population wo upzoning
English
0
0
0
6
Joseph Brown
Joseph Brown@heresyfinancial·
Whoever is in charge of TV prices should be put in charge of healthcare, education, and housing prices
Joseph Brown tweet media
English
2K
838
6.1K
6M
John Macgowan
John Macgowan@john_macgowan·
@Ben_Thompson247 I don't know if non citizens are eligible for HECs, but until yesterday I thought international students bringing their families out was just a meme so who knows, nothing surprises at this point
English
1
0
4
148
John Macgowan
John Macgowan@john_macgowan·
You could slash the price of HECs forgiveness substantially and target it only to the dip shits who failed to calculate there wouldn't be an employment market for basket weavers by saying if you take the reduction - you can't apply for HECs again.
English
9
4
69
4.4K
Pineapple Crab
Pineapple Crab@AdenGlaven1994·
The problem with the AUS office is that white collar comedy has been done to death. What we need is a show called “The Warehouse”. The ethnically diverse characters of a coffee distribution house in North West Melbourne featuring everyday convos and workplace politics.
English
32
46
917
27.9K
apenwarr
apenwarr@apenwarr·
Are there GPU Cloud Arbitrage companies, where you pay them a fee and they find you the cheapest cloud GPUs of the day to run your stuff on? I can’t believe every company is doing this by hand. (Or can I?)
English
8
2
40
5.8K
Andrew Garrett
Andrew Garrett@werdnum·
@johnb78 I have a half baked theory that there are 3 kinds of goods: those that benefit from productivity gains (most material goods, prices trend to zero), those that don't (labour intensive services), and "rare" goods (land, celebrity access) that don't scale.
English
0
0
1
46