The concept of Product-Market Fit (PMF) collapse has gained renewed attention with the rise of large language models (LLMs), as highlighted in a recent Reforge article. The article argues we’re witnessing unprecedented market disruption, in this post, I propose we’re experiencing an acceleration of a familiar pattern rather than a fundamentally new phenomenon. Adoption Curves […] The post The Exodus Curve appeared first on Marc Astbury.

On a few photowalks around Kitsilano in Vancouver BC with my first roll of Fomapan Classic 100. All pictures shot with older, non-AI Nikkor 50mm f/1.4 lens. Developed in Ilford Ilfosol3 for 5 minutes at 20°C and scanned with a cheap film scanner. Color has been removed and some levels adjusted slightly. Gate to the… Continue reading 5 Frames with Nikon FM and Fomapan 100

In 1940, President Roosevelt tapped William S. Knudsen to run the government's production of military equipment. Knudsen had spent a pivotal decade at Ford during the mass-production revolution, and was president of General Motors, when he was drafted as a civilian into service as a three-star general. Not bad for a Dane, born just ten minutes on bike from where I'm writing this in Copenhagen! Knudsen's leadership raised the productive capacity of the US war machine by a 100x in areas like plane production, where it went from producing 3,000 planes in 1939 to over 300,000 by 1945. He was quoted on his achievement: "We won because we smothered the enemy in an avalanche of production, the like of which he had never seen, nor dreamed possible". Knudsen wasn't an elected politician. He wasn't even a military man. But Roosevelt saw that this remarkable Dane had the skills needed to reform a puny war effort into one capable of winning the Second World War. Do you see where I'm going with this? Elon Musk is a modern day William S. Knudsen. Only even more accomplished in efficiency management, factory optimization, and first-order systems thinking. No, America isn't in a hot war with the Axis powers, but for the sake of the West, it damn well better be prepared for one in the future. Or better still, be so formidable that no other country or alliance would even think to start one. And this requires a strong, confident, and sound state with its affairs in order. If you look at the government budget alone, this is direly not so. The US was knocking on a two-trillion-dollar budget deficit in 2024! Adding to a towering debt that's now north of 36 trillion. A burden that's already consuming $881 billion in yearly interest payments. More than what's spent on the military or Medicare. Second to only Social Security on the list of line items. Clearly, this is not sustainable. This is the context of DOGE. The program, lead by Musk, that's been deputized by Trump to turn the ship around. History doesn't repeat, but it rhymes, and Musk is dropping beats that Knudsen would have surely been tapping his foot to. And just like Knudsen in his time, it's hard to think of any other American entrepreneur more qualified to tackle exactly this two-trillion dollar problem.  It is through The Musk Algorithm that SpaceX lowered the cost of sending a kilo of goods into lower orbit from the US by well over a magnitude. And now America's share of worldwide space transit has risen from less than 30% in 2010 to about 85%. Thanks to reusable rockets and chopstick-catching landing towers. Thanks to Musk. Or to take a more earthly example with Twitter. Before Musk took over, Twitter had revenues of $5 billion and earned $682 million. After the take over, X has managed to earn $1.25 billion on $2.7 billion in revenue. Mostly thank to the fact that Musk cut 80% of the staff out of the operation, and savaged the cloud costs of running the service. This is not what people expected at the time of the take over! Not only did many commentators believe that Twitter was going to collapse from the drastic costs in staff, they also thought that the financing for the deal would implode. Chiefly as a result of advertisers withdrawing from the platform under intense media pressure. But that just didn't happen. Today, the debt used to take over Twitter and turn it into X is trading at 97 cents on the dollar. The business is twice as profitable as it was before, and arguably as influential as ever. All with just a fifth of the staff required to run it. Whatever you think of Musk and his personal tweets, it's impossible to deny what an insane achievement of efficiency this has been! These are just two examples of Musk's incredible ability to defy the odds and deliver the most unbelievable efficiency gains known to modern business records. And we haven't even talked about taking Tesla from producing 35,000 cars in 2014 to making 1.7 million in 2024. Or turning xAI into a major force in AI by assembling a 100,000 H100 cluster at "superhuman" pace.  Who wouldn't want such a capacity involved in finding the waste, sloth, and squander in the US budget? Well, his political enemies, of course! And I get it. Musk's magic is balanced with mania and even a dash of madness. This is usually the case with truly extraordinary humans. The taller they stand, the longer the shadow. Expecting Musk to do what he does and then also be a "normal, chill dude" is delusional. But even so, I think it's completely fair to be put off by his tendency to fire tweets from the hip, opine on world affairs during all hours of the day, and offer his support to fringe characters in politics, business, and technology. I'd be surprised if even the most ardent Musk super fans don't wince a little every now and then at some of the antics. And yet, I don't have any trouble weighing those antics against the contributions he's made to mankind, and finding an easy and overwhelming balance in favor of his positive achievements. Musk is exactly the kind of formidable player you want on your team when you're down two trillion to nothing, needing a Hail Mary pass for the destiny of America, and eager to see the West win the future. He's a modern-day Knudsen on steroids (or Ketamine?). Let him cook.

Last week, James Truitt asked a question on Mastodon: James Truitt (he/him) @linguistory@code4lib.social Mastodon #digipres folks happen to have a handy repo of small invalid bags for testing purposes? I'm trying to automate our ingest process, and want to make sure I'm accounting for as many broken expectations as possible. Jan 31, 2025 at 07:49 PM The “bags” he’s referring to are BagIt bags. BagIt is an open format developed by the Library of Congress for packaging digital files. Bags include manifests and checksums that describe their contents, and they’re often used by libraries and archives to organise files before transfering them to permanent storage. Although I don’t use BagIt any more, I spent a lot of time working with it when I was a software developer at Wellcome Collection. We used BagIt as the packaging format for files saved to our cloud storage service, and we built a microservice very similar to what James is describing. The “bag verifier” would look for broken bags, and reject them before they were copied to long-term storage. I wrote a lot of bag verifier test cases to confirm that it would spot invalid or broken bags, and that it would give a useful error message when it did. All of the code for Wellcome’s storage service is shared on GitHub under an MIT license, including the bag verifier tests. They’re wrapped in a Scala test framework that might not be the easiest thing to read, so I’m going to describe the test cases in a more human-friendly way. Before diving into specific examples, it’s worth remembering: context is king. BagIt is described by RFC 8493, and you could create invalid bags by doing a line-by-line reading and deliberately ignoring every “MUST” or “SHOULD” but I wouldn’t recommend this aproach. You’d get a long list of test cases, but you’d be overwhelmed by examples, and you might miss specific requirements for your system. The BagIt RFC is written for the most general case, but if you’re actually building a storage service, you’ll have more concrete requirements and context. It’s helpful to look at that context, and how it affects the data you want to store. Who’s creating the bags? How will they name files? Where are you going to store bags? How do bags fit into your wider systems? And so on. Understanding your context will allow you to skip verification steps that you don’t need, and to add verification steps that are important to you. I doubt any two systems implement the exact same set of checks, because every system has different context. Here are examples of potential validation issues drawn from the BagIt specification and my real-world experience. You won’t need to check for everything on this list, and this list isn’t exhaustive – but it should help you think about bag validation in your own context. The Bag Declaration bagit.txt This file declares that this is a BagIt bag, and the version of BagIt you’re using (RFC 8493 §2.1.1). It looks the same in almost every bag, for example: BagIt-Version: 1.0 Tag-File-Character-Encoding: UTF-8 This tightly prescribed format means it can only be invalid in a few ways: What if the bag doesn’t have a bag declaration? It’s a required element of every BagIt bag; it has to be there. What if the bag declaration is the wrong format? It should contain exactly two lines: a version number and a character encoding, in that order. What if the bag declaration has an unexpected version number? If you see a BagIt version that you’ve not seen before, the bag might have a different structure than what you expect. The Payload Files and Payload Manifest The payload files are the actual content you want to save and preserve. They get saved in the payload directory data/ (RFC 8493 §2.1.2), and there’s a payload manifest manifest-algorithm.txt that lists them, along with their checksums (RFC 8493 §2.1.3). Here’s an example of a payload manifest with MD5 checksums: 37d0b74d5300cf839f706f70590194c3 data/waterfall.jpg This tells us that the bag contains a single file data/waterfall.jpg, and it has the MD5 checksum 37d0…. These checksums can be used to verify that the files have transferred correctly, and haven’t been corrupted in the process. There are lots of ways a payload manifest could be invalid: What if the bag doesn’t have a payload manifest? Every BagIt bag must have at least one Payload Manifest file. What if the payload manifest is the wrong format? These files have a prescribed format – one file per line, with a checksum and file path. What if the payload manifest refers to a file that isn’t in the bag? Either one of the files in the bag has been deleted, or the manifest has an erroneous entry. What if the bag has a file that isn’t listed in the payload manifest? The manifest should be a complete listing of all the payload files in the bag. If the bag has a file which isn’t in the payload manifest, either that file isn’t meant to be there, or the manifest is missing an entry. Checking for unlisted files is how I spotted unwanted .DS_Store and Thumbs.db files. What if the checksum in the payload manifest doesn’t match the checksum of the file? Either the file has been corrupted, or the checksum is incorrect. What if there are payload files outside the data/ directory? All the payload files should be stored in data/. Anything outside that is an error. What if there are duplicate entries in the payload manifest? Every payload file must be listed exactly once in the manifest. This avoids ambiguity – suppose a file is listed twice, with two different checksums. Is the bag valid if one of those checksums is correct? Requiring unique entries avoids this sort of issue. What if the payload directory is empty? This is perfectly acceptable in the BagIt RFC, but it may not be what you want. If you know that you will always be sending bags that contain files, you should flag empty payload directories as an error. What if the payload manifest contains paths outside data/, or relative paths that try to escape the bag? (e.g. ../file.txt) Now we’re into “malicious bag” territory – a bag uploaded by somebody who’s trying to compromise your ingest pipeline. Any such bags should be treated with suspicion and rejected. If you’re concerned about malicious bags, you need a more thorough test suite to catch other shenanigans. We never went this far at Wellcome Collection, because we didn’t ingest bags from arbitrary sources. The bags only came from internal systems, and our verification was mainly about spotting bugs in those systems, not defending against malicious actors. A bag can contain multiple payload manifests – for example, it might contain both MD5 and SHA1 checksums. Every payload manifest must be valid for the overall bag to be valid. Payload filenames There are lots of gotchas around filenames and paths. It’s a complicated problem, and I definitely don’t understand all of it. It’s worth understanding the filename rules of any filesystem where you will be storing bags. For example, Azure Blob Storage has a number of rules around how you can name files, and Amazon S3 has different rules. We stored files in both at Wellcome Collection, and so the storage service had to enforce the superset of these rules. I’ve listed some edge cases of filenames you might want to consider, but it’s not a comlpete list. There are lots of ways that unexpected filenames could cause you issues, but whether you care depends on the source of your bags. If you control the bags and you know you’re not going to include any weird filenames, you can probably skip most of these. We only checked for one of these conditions at Wellcome Collection, because we had a pre-ingest step that normalised filenames. It converted filenames to ASCII, and saved a mapping between original and normalised filename in the bag. However, the normalisation was only designed for one filesystem, and produced filenames with trailing dots that were still disallowed in Azure Blob. What if a filename is too long? Some systems have a maximum path length, and an excessively deep directory structure or long filename could cause issues. What if a filename contains special characters? Spaces, emoji, or special characters (\, :, *, etc.) can cause problems for some tools. You should also think about characters that need to be URL-encoded. What if a filename has trailing spaces or dots? Some filesystems can’t support filenames ending in a dot or a space. What happens if your bag contains such a file, and you try to save it to the filesystem? This caused us issues at Wellcome Collection. We initially stored bags just in Amazon S3, which is happy to take filenames with a trailing dot – then we added backups to Azure Blob, which doesn’t. One of the bags we’d stored in Amazon S3 had a trailing dot in the filename, and caused us headaches when we tried to copy it to Azure. What if a filename contains a mix of path separators? The payload manifest uses a forward slash (/) as a path separator. If you have a filename with an alternative path separator, it might behave differently on different systems. For example, consider the payload file a\b\c. This would be a single file on macOS or Linux, but it would be nested inside two folders on Windows. What if the filenames are a mix of uppercase and lowercase characters? Some fileystems are case-sensitive, others aren’t. This can cause issues when you move bags between systems. For example, suppose a bag contains two different files Macrodata.txt and macrodata.txt. When you save that bag on a case-insensitive filesystem, only one file will be saved. What if the same filename appears twice with different Unicode normalisations? This is similar to filenames which only differ in upper/lowercase. They might be treated as two files on one filesystem, but collapsed into one file on another. The classic example is the word “café”: this can be encoded as caf\xc3\xa9 (UTF-8 encoded é) or cafe\xcc\x81 (e + combining acute accent). What if a filename contains a directory reference? A directory reference is /./ (current directory) or /../ (parent directory). It’s used on both Unix and Windows-like systems, and it’s another case of two filenames that look different but can resolve to the same path. For example: a/b, a/./b and a/subdir/../b all resolve to the same path under these rules. This can cause particular issues if you’re moving between local filesystems and cloud storage. Local filesystems treat filenames as hierarchical paths, where cloud storage like Amazon S3 often treats them as opaque strings. This can cause issues if you try to copy files from cloud storage to a local system – if you’re not careful, you could lose files in the process. The Tag Manifest tagmanifest-algorithm.txt Similar to the payload manifest, the tag manifest lists the tag files and their checksums. A “tag file” is the BagIt term for any metadata file that isn’t part of the payload (RFC 8493 §2.2.1). Unlike the payload manifest, the tag manifest is optional. A bag without a tag manifest can still be a valid bag. If the tag manifest is present, then many of the ways that a payload manifest can invalidate a bag – malformed contents, unreferenced files, or incorrect checksums – can also apply to tag manifests. There are some additional things to consider: What if a tag manifest lists payload files? The tag manifest lists tag files; the payload manifest lists payload files in the data/ directory. A tag manifest that lists files in the data/ directory is incorrect. What if the bag has a file that isn’t listed in either manifest? Every file in a bag (except the tag manifests) should be listed in either a payload or a tag manifest. A file that appears in neither could mean an unexpected file, or a missing manifest entry. Although the tag manifest is optional in the BagIt spec, at Wellcome Collection we made it a required file. Every bag had to have at least one tag manifest file, or our storage service would refuse to ingest it. The Bag Metadata bag-info.txt This is an optional metadata file that describes the bag and its contents (RFC 8493 §2.2.2). It’s a list of metadata elements, as simple label-value pairs, one per line. Here’s an example of a bag metadata file: Source-Organization: Lumon Industries Organization-Address: 100 Main Street, Kier, PE, 07043 Contact-Name: Harmony Cobel Unlike the manifest files, this is primarily intended for human readers. You can put arbitrary metadata in here, so you can add fields specific to your organisation. Although this file is more flexible, there are still ways it can be invalid: What if the bag metadata is the wrong format? It should have one metadata entry per line, with a label-value pair that’s separated by a colon. What if the Payload-Oxum is incorrect? The Payload-Oxum contains some concise statistics about the payload files: their total size in bytes, and how many there are. For example: Payload-Oxum: 517114.42 This tells us that the bag contains 42 payload files, and their total size is 517,114 bytes. If these stats don’t match the rest of the bag, something is wrong. What if non-repeatable metadata element names are repeated? The BagIt RFC defines a small number of reserved metadata element names which have a standard meaning. Although most metadata element names can be repeated, there are some which can’t, because they can only have one value. In particular: Bagging-Date, Bag-Size, Payload-Oxum and Bag-Group-Identifier. Although the bag metadata file is optional in a general BagIt bag, you may want to add your own rules based on how you use it. For example, at Wellcome Collection, we required all bags to have an External-Identifier value, that matched a specific schema. This allowed us to link bags to records in other databases, and our bag verifier would reject bags that didn’t include it. The Fetch File fetch.txt This is an optional element that allows you to reference files stored elsewhere (RFC 8493 §2.2.3). It tells the person reading the bag that a file hasn’t been included in this copy of the bag; they have to go and fetch it from somewhere else. The file is still recorded in the payload manifest (with a checksum you can verify), but you don’t have a complete bag until you’ve downloaded all the files. Here’s an example of a fetch.txt: https://topekastar.com/~daria/article.txt 1841 data/article.txt This tells us that data/article.txt isn’t included in this copy of the bag, but we we can download it from https://topekastar.com/~daria/article.txt. (The number 1841 is the size of the file in bytes. It’s optional.) Using fetch.txt allows you to send a bag with “holes”, which saves disk space and network bandwidth, but at a cost – we’re now relying on the remote location to remain available. From a preservation standpoint, this is scary! If topekastar.com goes away, this bag will be broken. I know some people don’t use fetch.txt for precisely this reason. If you do use fetch.txt, here are some things to consider: What if the fetch file is the wrong format? There’s a prescribed format – one file per line, with a URL, optional file size, and file path. What if the fetch file lists a file which isn’t in the payload manifest? The fetch.txt should only tell us that a file is stored elsewhere, and shouldn’t be introducing otherwise unreferenced files. If a file appears in fetch.txt but not the payload manifest, then we can’t verify the remote file because we don’t have a checksum for it. There’s either an erroneous fetch file entry or a missing manifest entry. What if the fetch file points to a file at an unusable URL? The URL is only useful if the person who receives the bag can use it to download the file. If they can’t, the bag might technically be valid, but it’s functionally broken. For example, you might reject URLs that don’t start with http:// or https://. What if the fetch file points to a file with the wrong length? The fetch.txt can optionally specify the size of a file, so you know how much storage you need to download it. If you download the file, the actual size should match the stated size. What if the fetch files points to a file that’s already included in the bag? Now you have two ways to get this file: you can read it from the bag, or from the remote URL. If a file is listed in both fetch.txt and included in the bag, either that file isn’t meant to be in the bag, or the fetch file has an erroneous entry. We used fetch files at Wellcome Collection to implement versioning, and we added extra rules about what remote URLs were allowed. In particular, we didn’t allow fetching a file from just anywhere – you could fetch from our S3 buckets, but not the general Internet. The bag verifier would reject a fetch file entry that pointed elsewhere. These examples illustrate just how many ways a BagIt bag can be invalid, from simple structural issues to complex edge cases. Remember: the key is to understand your specific needs and requirements. By considering your context – who creates your bags, where they’ll be stored, and how they fit into your wider systems – you can build a validation process to catch the issues that matter to you, while avoiding unnecessary complexity. I can give you my ideas, but only you can build your system. [If the formatting of this post looks odd in your feed reader, visit the original article]

Quoting myself from former days on Twitter: Businesses have a mental model of what they do. Businesses build software to help them do it—a concrete manifestation of their mental model. A gap always exists between these two. What makes a great software business is their ability to keep that gap very small. I think this holds up. And I still think about this idea (hence this post). Software is an implementation of human understanding — people need X, so we made Y. But people change. Businesses change. So software must also change. One of your greatest strengths will be your ability to adapt and evolve your understanding of people’s needs and implement it in your software. In a sense, technical debt is the other side of this coin of change: an inability to keep up with your own metamorphosis and understanding. In a way, you could analogize this to the conundrum of rocket science: you need fuel to get to space, but the more fuel you add, the more weight you add, and the more weight you add, the more fuel you need. Ad nauseam. It’s akin to making software. You want to make great software for people’s needs today. It takes people, processes, and tools to make software, but the more people, processes, and tools you add to the machine of making software, the less agile you become. So to gain velocity you add more people, processes, and tools, which…you get the idea. Being able to build and maintain pliable software that can change and evolve at the same speed as your mental model is a superpower. Quality in code means the flexibility to change. Email :: Mastodon :: Bluesky