I was expecting a white Christmas. Have not had this for a long time. So surely Iceland will have one, right? Well, turns out that we had a lot of wind here (20m/s) and a bit of rain, but also plus 8°C. Definitely not white here in the west of Iceland.

Eine einfache und leckere Kartoffelsuppe: https://emmikochteinfach.de/omas-kartoffelsuppe-mit-wuerstchen/

Änderungen die ich gemacht habe:

• doppelt soviel Knollensellerie
• Keine Petersilienwurzel
• 200g Zwiebel
• keine Créme Fraîche
• kein Speck, sondern Salami (hatte nix anderes)
• kein Knoblauch
• kein Senf
• keine Petersilie

Schmeckt trotzdem!

I don’t like the AI generated answer/summaries from Google. I can read myself, and I actually prefer to read the original information to get an idea whether it’s coherent and correct (as much as I can say since usually I look for information for which I do not know the answer).

An AI summary gives me a coherent text with unknown degrees of truth. But it always reads like it was written by someone who can make a coherent sentence.

There’s a “hack” to add some swear words which removed the AI summary, but it looks like a clunky workaround. Maybe 10 year olds get a kick out of adding “fuck” into every Google search, but I am too old for that shit.

This is a good solution though: https://udm14.com/

No more AI summary. Feels like the old and useful Google. Love it!

To make the default search engine use this switch, follow the instructions at https://tenbluelinks.org/ and your normal Google search will be using that udm=14 parameter.

Of course you could also use https://duckduckgo.com/ as your default search engine…

I find a lot of LLM uses which are anywhere from pointless to futile, but once in a while I find an example of LLM use which adds value and is something which would have been much harder without LLM.

Accelerating Large-Scale Test Migration with LLMs from AirBnB is one of those:

• Tests are short snippets of code which don’t contain complex logic: they do set up a test, run it, and check the result.
• Tests are supposed to pass, thus it’s easy and quick to verify the results created by the LLM.
• LLMs have seen enough JavaScript to know how to create valid code. Same for the two involved testing frameworks: Enzyme and the React Testing Library.
• If the generated tests don’t run at all (i.e. syntax errors or severe logical errors), then just try again. Not much harm done.
• Adding context works, thus use an LLM which can handle a large enough context window. And make sure your context is relevant and helpful. In above examples they went up to 100k token input.

Reading above list, it’s easy to imagine where LLM will be much less successful:

• Unusual languages or frameworks
• Hard to verify that the LLM did good

Costs

I wish the authors would have mentioned the costs for this migration. I am confident it was way cheaper than doing it manually. They estimated 1.5 years vs. 6 weeks it took with LLM.

From the article, I am assuming 3600 files in total and 75% were solved with a prompt with 50k token on average, assuming 5 tries and using Claude 3.7 Sonnet at $3/1M input token, and ignoring the output token count since it’s relatively small code, that’s $2000 for the 2700 files which worked, and $1350 for the 900 which failed even after 10 tries.

Add in the attempt to fix those 900 remaining files with different prompts, adding different examples or context files etc. and the 50-100 times rerunning the LLM prompt, that’s another $6000. And then manually fixing the few (3%, about 100 files) remaining files “in another week of work”.

So roughly $10,000 for the LLM and 1.5 months for a small team. Compared to 18 months for a relatively simple programming task. Since tests are independent from one another, that could be done in 1.5 months by hiring 12 people instead of 1. Same price but 12 times faster.

I work in a data center and we replace a ton of SSD. And those are the good ones, not the cheaper consumer models. Thus I know SSDs can break.

Long time ago I had a HDD die on me (Seagate ST 1162N, 142MB, SCSI). The model and the capacity gives away that this is long ago. I learned my lesson and since then backups are semi-regularly done. My NAS has RAID1 too. No severe data loss since then.

Ignoring microSD cards, which I do not count, I had yet to see one of my SSD failing. Today I got one:

[  145.916941] sd 1:0:0:0: [sda] tag#0 FAILED Result: hostbyte=DID_OK driverbyte=DRIVER_OK cmd_age=0s
[  145.916948] sd 1:0:0:0: [sda] tag#0 Sense Key : Illegal Request [current] 
[  145.916954] sd 1:0:0:0: [sda] tag#0 Add. Sense: Invalid command operation code
[  145.916960] sd 1:0:0:0: [sda] tag#0 CDB: Write(10) 2a 00 06 ce f0 38 00 00 18 00
[  145.916963] critical target error, dev sda, sector 114225208 op 0x1:(WRITE) flags 0x800 phys_seg 3 prio class 2
[  147.061324] JBD2: recovery failed
[  147.061333] EXT4-fs (dm-7): error loading journal

The culprit was a Klevv Cras C700 NVMe SSD with 480GB and Silicon Motion SM2263EN controller. Worked well. Until it stopped working.

smartctl shows nothing odd:

root@m75q:~# smartctl --all /dev/sda
smartctl 7.3 2022-02-28 r5338 [x86_64-linux-6.1.0-32-amd64] (local build)
Copyright (C) 2002-22, Bruce Allen, Christian Franke, www.smartmontools.org

=== START OF INFORMATION SECTION ===
Model Number:                       KLEVV CRAS C700 M.2 NVMe SSD 480GB
Serial Number:                      E201908210020546
Firmware Version:                   R0801L2
PCI Vendor/Subsystem ID:            0x126f
IEEE OUI Identifier:                0x005cd2
Controller ID:                      1
NVMe Version:                       1.3
Number of Namespaces:               1
Namespace 1 Size/Capacity:          480,103,981,056 [480 GB]
Namespace 1 Formatted LBA Size:     512
Namespace 1 IEEE EUI-64:            000000 0000000028
Local Time is:                      Sun Mar 23 22:04:58 2025 JST
Firmware Updates (0x14):            2 Slots, no Reset required
Optional Admin Commands (0x0006):   Format Frmw_DL
Optional NVM Commands (0x005f):     Comp Wr_Unc DS_Mngmt Wr_Zero Sav/Sel_Feat Timestmp
Log Page Attributes (0x0f):         S/H_per_NS Cmd_Eff_Lg Ext_Get_Lg Telmtry_Lg
Maximum Data Transfer Size:         32 Pages
Warning  Comp. Temp. Threshold:     70 Celsius
Critical Comp. Temp. Threshold:     80 Celsius

Supported Power States
St Op     Max   Active     Idle   RL RT WL WT  Ent_Lat  Ex_Lat
 0 +     9.00W       -        -    0  0  0  0        0       0
 1 +     4.60W       -        -    1  1  1  1        0       0
 2 +     3.80W       -        -    2  2  2  2        0       0
 3 -   0.0450W       -        -    3  3  3  3     2000    2000
 4 -     0.01W       -        -    4  4  4  4     8000 1015936

Supported LBA Sizes (NSID 0x1)
Id Fmt  Data  Metadt  Rel_Perf
 0 +     512       0         0
 1 -    4096       0         0

=== START OF SMART DATA SECTION ===
SMART overall-health self-assessment test result: PASSED

SMART/Health Information (NVMe Log 0x02)
Critical Warning:                   0x00
Temperature:                        37 Celsius
Available Spare:                    100%
Available Spare Threshold:          10%
Percentage Used:                    0%
Data Units Read:                    3,035,689 [1.55 TB]
Data Units Written:                 10,340,813 [5.29 TB]
Host Read Commands:                 62,049,267
Host Write Commands:                193,235,592
Controller Busy Time:               4,569
Power Cycles:                       2,806
Power On Hours:                     6,365
Unsafe Shutdowns:                   29
Media and Data Integrity Errors:    123
Error Information Log Entries:      0
Warning  Comp. Temperature Time:    0
Critical Comp. Temperature Time:    0

Warning: NVMe Get Log truncated to 0x200 bytes, 0x200 bytes zero filled
Error Information (NVMe Log 0x01, 16 of 256 entries)
No Errors Logged

Interesting info from above:

• There are no errors logged.
• There are 123 media and data integrity errors.
• While I used that SSD for years and software development, I only wrote 5.29TB. That seems to be low, but it might be right.
• Reading only 1.55TB…that seems very low. It makes me wonder how correct this and all the other reported numbers are.
• 100% spare blocks exist. Why were they not used before it was too late?
• The computer was not turned on for almost a year. Maybe that’s the problem. Which would be a problem when using SSDs as offline backup storage. Makes a good case for HDDs for long term offline storage. Optical media makes even more sense.

Conclusion

• Have backups. Everything including SSDs can fatally fail suddenly.
• Don’t trust offline SSDs. Like every backup, test it regularly that it’s still good.
• I don’t have HDDs except in my NAS. I was planning to replace it with SSDs, but I’ll have to re-think that.

From the team “Keeping the Random in Harald’s Random Stuff”

I am lucky to have a Gyomu Super close to home: they are a normal super market with a lot of unusual things which are usually catering to professional food related businesses, e.g. restaurants. E.g. want 1kg of roasted onions? Or 2.5kg of canned whole Italian tomatoes?

They also have normal sizes, like 400g chocolate spread or marmalade. Not everything is super-sized.

If you order those, like actual businesses would do, you’ll get multiple of those delivered to you. Like a box with 12 packets of onions. Or 6 of those tomato cans. Or 12 glasses of chocolate spread.

However in their public supermarkets they sell those in single units. I love to go there once in a while if only to see what interesting things they currently sell.

Recently I found something I didn’t know existed: Panna Cotta or Chocolate Babaloa in an unusual shape of the typical 1l milk packs:

This is not a drink, but it’s what it says on the outside: 1kg pudding. In the shape of 1l milk packs. Shape aside, this stuff is so good and not even very expensive (500 Yen for the Babaloa), especially considering the quality. The is also Japanese Pudding (プリン), Coffee Jelly and some more.

If you have a sweet tooth and you are able to go to a Gyomu Super, get one! And get a frozen cheese cake or lemon cake too:

Those are a nice treat for weekends when you’d like a cake for “Kaffee und Kuchen“.