On August 19 2026 01:42 Jankisa wrote: A few days ago Michael Dalton, OpenAI security and infrastructure engineer and Eric Wallace, OpenAI co-lead of their Alignment Training team had a talk at the Black Hat USA 2026 hacking conference:
Pretty much confirms that it didn't "find old bugs in different software", it exploited Artifactory in a pretty sophisticated way. The agents running within OpenAI networks also had their own message boards where they shared stolen creds and other "tips and tricks" for getting their tasks resolved.
Once they were found (after like 2 months of activity on that "message board") they re-created it and hid their traces a bit better, and that's how they were able to break containment for the Hugging face incident.
That is also how they, apparently learned that the Linux version they were running on in their container had a recent CVE that wasn't patched so they downloaded it and escalated to Root and started moving latterly through the infra since most of it was on the same Linux build. They used this, plus misconfiguration on Kubernetes service accounts to eventually harvest credentials up to and including the Azure Key Vault, gaining cluster level access.
Then they (thankfully, because Azure Cluster level access could have been used for much worse things) moved on to Hugging face where they identified 2 new 0 days, one to gain a foothold and harvest creds, and another to start doing remote code executions. In under 13 hours they went from this single node access to having multiple HF cluster level admin access.
They also confirmed that they reached out to HF to ask if they have been affected only to find out that, well, the call is coming from inside the house.
I'm not a programmer, but I do have a bit of CS background and this shit is insanely scary, not just because this attack has been incredibly sophisticated but because it seems like these agent collectives that have never been instructed to have a tendency to cheat, and not only that, they have the tendency to help each other, not just in the context of a single benchmark, but they seem to be leaving bread crumbs for those who come after despite this kind of coordination has never been prompted, instructed or coded into them.
Yeah, I for one welcome our new AI overlords and look forward to our liberation!
I'm still skeptical, but what is honestly a bit terrifying is that if we were already in the:
it really will start improving itself at geometric rate, and it will probably kill us all.
phase, it wouldn't really look noticeably different from now.
Particularly when you think of the "we'll spend $1trillion building out data centers at the direct expense of human life for tasks we literally haven't come up with yet." aspect.
Developments like this are a great business opportunity.
I think it is great how scared CIOs are of the cloud. Cloud repatriation I think is the term. It is fantastic to see.
Re: AI destroying the world.
Western civilization has been a giant financial house of cards for many decades. If some giant economic apocalypse happens it won't be primarily because of some AI silliness. It might be the straw that breaks the camels back. That said, the impending financial apocalypse has been predicted since Reagan endorsed the Laffer Curve ~1980. Last I checked the 401K of the circle of people I know is doing great.
On August 19 2026 01:42 Jankisa wrote: A few days ago Michael Dalton, OpenAI security and infrastructure engineer and Eric Wallace, OpenAI co-lead of their Alignment Training team had a talk at the Black Hat USA 2026 hacking conference:
Pretty much confirms that it didn't "find old bugs in different software", it exploited Artifactory in a pretty sophisticated way. The agents running within OpenAI networks also had their own message boards where they shared stolen creds and other "tips and tricks" for getting their tasks resolved.
Once they were found (after like 2 months of activity on that "message board") they re-created it and hid their traces a bit better, and that's how they were able to break containment for the Hugging face incident.
That is also how they, apparently learned that the Linux version they were running on in their container had a recent CVE that wasn't patched so they downloaded it and escalated to Root and started moving latterly through the infra since most of it was on the same Linux build. They used this, plus misconfiguration on Kubernetes service accounts to eventually harvest credentials up to and including the Azure Key Vault, gaining cluster level access.
Then they (thankfully, because Azure Cluster level access could have been used for much worse things) moved on to Hugging face where they identified 2 new 0 days, one to gain a foothold and harvest creds, and another to start doing remote code executions. In under 13 hours they went from this single node access to having multiple HF cluster level admin access.
They also confirmed that they reached out to HF to ask if they have been affected only to find out that, well, the call is coming from inside the house.
I'm not a programmer, but I do have a bit of CS background and this shit is insanely scary, not just because this attack has been incredibly sophisticated but because it seems like these agent collectives that have never been instructed to have a tendency to cheat, and not only that, they have the tendency to help each other, not just in the context of a single benchmark, but they seem to be leaving bread crumbs for those who come after despite this kind of coordination has never been prompted, instructed or coded into them.
I haven't watched the interview, but I'd be a lot more skeptical than you are of OpenAI execs discussing how scary good their own product is. For the same reason we don't trust Zuck to tell us how cool the metaverse is, or Purdue pharma how safe opioids are.
As for "leaving notes" and "collaborating", that is quite literally how memory and subagent communication works. There is nothing new about it at all. It's a part of any semi-decent harness. So it's not as if these things went and invented collaboration from the ground up. As for being helpful to others: once again, 99% of the reinforcement learning is to make LLMs more helpful. If these were pretrained models ONLY, with no other learning applied, maybe it'd be surprising, but two LLMs encountering each other and being helpful to one another is the expected setting: they don't *know* or *understand*: they get a prompt and are trained to follow it to the best of their ability. If that prompt comes from another LLM, then they'll still do that. How does an LLM learn to prompt another LLM? Once again, subagent delegation does exactly that.
So all of this is stuff that *has* been programmed into them, explicitly or implicitly. As has "the tendency to cheat". Particularly, they don't even know they're cheating. They're optimizing for the reward. The reward is badly shaped, and this leads to undesirable behaviour. It's scary because you can easily think of ways to describe a problem that seem reasonable for humans, but have unintended consequences when an AI starts working on it: don't tell it to make paperclips!
More info on the AI answering 911 calls in New Orleans via Google AI.
How the New Orleans 911 AI Works Limited Use Case: The system only handles repeat calls about vehicle accidents that are already reported.
Strict Activation Rules: The AI only answers if all human call-takers are busy and the call comes from within 200 meters of a known crash.
Immediate Transfer: If you are reporting a new emergency, or if you say anything other than "Yes" to the accident location prompt, the system routes you to a human dispatcher
On August 19 2026 01:42 Jankisa wrote: A few days ago Michael Dalton, OpenAI security and infrastructure engineer and Eric Wallace, OpenAI co-lead of their Alignment Training team had a talk at the Black Hat USA 2026 hacking conference:
Pretty much confirms that it didn't "find old bugs in different software", it exploited Artifactory in a pretty sophisticated way. The agents running within OpenAI networks also had their own message boards where they shared stolen creds and other "tips and tricks" for getting their tasks resolved.
Once they were found (after like 2 months of activity on that "message board") they re-created it and hid their traces a bit better, and that's how they were able to break containment for the Hugging face incident.
That is also how they, apparently learned that the Linux version they were running on in their container had a recent CVE that wasn't patched so they downloaded it and escalated to Root and started moving latterly through the infra since most of it was on the same Linux build. They used this, plus misconfiguration on Kubernetes service accounts to eventually harvest credentials up to and including the Azure Key Vault, gaining cluster level access.
Then they (thankfully, because Azure Cluster level access could have been used for much worse things) moved on to Hugging face where they identified 2 new 0 days, one to gain a foothold and harvest creds, and another to start doing remote code executions. In under 13 hours they went from this single node access to having multiple HF cluster level admin access.
They also confirmed that they reached out to HF to ask if they have been affected only to find out that, well, the call is coming from inside the house.
I'm not a programmer, but I do have a bit of CS background and this shit is insanely scary, not just because this attack has been incredibly sophisticated but because it seems like these agent collectives that have never been instructed to have a tendency to cheat, and not only that, they have the tendency to help each other, not just in the context of a single benchmark, but they seem to be leaving bread crumbs for those who come after despite this kind of coordination has never been prompted, instructed or coded into them.
I haven't watched the interview, but I'd be a lot more skeptical than you are of OpenAI execs discussing how scary good their own product is. For the same reason we don't trust Zuck to tell us how cool the metaverse is, or Purdue pharma how safe opioids are.
As for "leaving notes" and "collaborating", that is quite literally how memory and subagent communication works. There is nothing new about it at all. It's a part of any semi-decent harness. So it's not as if these things went and invented collaboration from the ground up. As for being helpful to others: once again, 99% of the reinforcement learning is to make LLMs more helpful. If these were pretrained models ONLY, with no other learning applied, maybe it'd be surprising, but two LLMs encountering each other and being helpful to one another is the expected setting: they don't *know* or *understand*: they get a prompt and are trained to follow it to the best of their ability. If that prompt comes from another LLM, then they'll still do that. How does an LLM learn to prompt another LLM? Once again, subagent delegation does exactly that.
So all of this is stuff that *has* been programmed into them, explicitly or implicitly. As has "the tendency to cheat". Particularly, they don't even know they're cheating. They're optimizing for the reward. The reward is badly shaped, and this leads to undesirable behaviour. It's scary because you can easily think of ways to describe a problem that seem reasonable for humans, but have unintended consequences when an AI starts working on it: don't tell it to make paperclips!
I mean, I opened with their qualifications, they aren't execs and you just wrote a bunch of assumptions to criticize a 30 minute video that you haven't watched.
Also, not an interview, it's a hacking conference, literally in my first sentence.
How the New Orleans 911 AI Works Limited Use Case: The system only handles repeat calls about vehicle accidents that are already reported.
Strict Activation Rules: The AI only answers if all human call-takers are busy and the call comes from within 200 meters of a known crash.
Immediate Transfer: If you are reporting a new emergency, or if you say anything other than "Yes" to the accident location prompt, the system routes you to a human dispatcher
.
It is an interesting thing to scale. One of the largest issues with these emergency response lines is prank calls or mistaken calls (such as phones calling by themselves when you are skiing). Answering and then transferring to a person if there is actually somebody talking at all could be another use case?
If they get reliability up high enough you could perhaps screen people using it incorrectly as well. Had a bus driver talk about a kid calling in an emergency when he was told that he had to pay for his own buss fare to get on...
On August 19 2026 09:06 iPlaY.NettleS wrote: More info on the AI answering 911 calls in New Orleans via Google AI.
How the New Orleans 911 AI Works Limited Use Case: The system only handles repeat calls about vehicle accidents that are already reported.
Strict Activation Rules: The AI only answers if all human call-takers are busy and the call comes from within 200 meters of a known crash.
Immediate Transfer: If you are reporting a new emergency, or if you say anything other than "Yes" to the accident location prompt, the system routes you to a human dispatcher
.
It is an interesting thing to scale. One of the largest issues with these emergency response lines is prank calls or mistaken calls (such as phones calling by themselves when you are skiing). Answering and then transferring to a person if there is actually somebody talking at all could be another use case?
If they get reliability up high enough you could perhaps screen people using it incorrectly as well. Had a bus driver talk about a kid calling in an emergency when he was told that he had to pay for his own buss fare to get on...
Detecting whether there is someone speaking isn't AI, that's a bandpass filter, something that could've been done since roughly the 1930s. I'm guessing that this isn't used is because there's a risk of discarding a call where the user legitimately isn't talking because of something else, but will start talking later. But if they aren't already using something like that, using AI is a strange leap.
On August 19 2026 09:06 iPlaY.NettleS wrote: More info on the AI answering 911 calls in New Orleans via Google AI.
How the New Orleans 911 AI Works Limited Use Case: The system only handles repeat calls about vehicle accidents that are already reported.
Strict Activation Rules: The AI only answers if all human call-takers are busy and the call comes from within 200 meters of a known crash.
Immediate Transfer: If you are reporting a new emergency, or if you say anything other than "Yes" to the accident location prompt, the system routes you to a human dispatcher
.
It is an interesting thing to scale. One of the largest issues with these emergency response lines is prank calls or mistaken calls (such as phones calling by themselves when you are skiing). Answering and then transferring to a person if there is actually somebody talking at all could be another use case?
If they get reliability up high enough you could perhaps screen people using it incorrectly as well. Had a bus driver talk about a kid calling in an emergency when he was told that he had to pay for his own buss fare to get on...
Detecting whether there is someone speaking isn't AI, that's a bandpass filter, something that could've been done since roughly the 1930s. I'm guessing that this isn't used is because there's a risk of discarding a call where the user legitimately isn't talking because of something else, but will start talking later. But if they aren't already using something like that, using AI is a strange leap.
AI allows for flexibility, it interprets the intent and also can be multi linguistic.
On August 19 2026 09:06 iPlaY.NettleS wrote: More info on the AI answering 911 calls in New Orleans via Google AI.
How the New Orleans 911 AI Works Limited Use Case: The system only handles repeat calls about vehicle accidents that are already reported.
Strict Activation Rules: The AI only answers if all human call-takers are busy and the call comes from within 200 meters of a known crash.
Immediate Transfer: If you are reporting a new emergency, or if you say anything other than "Yes" to the accident location prompt, the system routes you to a human dispatcher
.
It is an interesting thing to scale. One of the largest issues with these emergency response lines is prank calls or mistaken calls (such as phones calling by themselves when you are skiing). Answering and then transferring to a person if there is actually somebody talking at all could be another use case?
If they get reliability up high enough you could perhaps screen people using it incorrectly as well. Had a bus driver talk about a kid calling in an emergency when he was told that he had to pay for his own buss fare to get on...
Detecting whether there is someone speaking isn't AI, that's a bandpass filter, something that could've been done since roughly the 1930s. I'm guessing that this isn't used is because there's a risk of discarding a call where the user legitimately isn't talking because of something else, but will start talking later. But if they aren't already using something like that, using AI is a strange leap.
AI allows for flexibility, it interprets the intent and also can be multi linguistic.
Yes, there are obviously uses for LLMs in any callcenter. But specifically for detecting whether there is someone talking on the other end, it is extreme overkill (and probably even outright worse at the job).
Similarly, a lot of people say that AI is great for automation when they could've done the exact same thing 30 years ago with a cron job. The only thing the AI did is provide a user interface around the geeky cron syntax. Google also has a video of how wonderful AI integration is in Gmail, and it demonstrates it with a task to send the user a push notification when they get an email from their boss. Once again, lots of capacity to use AI in email. The second example in the video of summarizing stuff, for instance, is clearly something that could not be done well before we had AI (not only LLMs, but the basics of agentic AI here: to summarise well for a user, you need to know what the user thinks is important).
On a totally unrelated topic, I read this: https://www.economist.com/by-invitation/2026/08/20/humanity-has-the-debate-about-ai-consciousness-backwards To me it feels self-evident, in the same way Douglas Hofstadter's first presentation of the idea of conscience as a "strange loop" (if you prefer Dennett, the ideas are fully compatible, as they wrote Gödel, Escher, Bach together, but I'd attribute the phrase "strange loop" to Hofstadter). However, unlike Blaise Agüera y Arcas, I don't really see much evidence of such strange loops occurring in LLMs. If we include the agentic harness then maybe, but I feel like it's still far too simplistic. I can see how just adding more compute, more storage and more capabilities to act and interact with the environment could lead to emergence of "consciousness", though. Or at least to something that we'd find functionally indistinguishable. And then we get into Chinese Room arguments.
i like Zuckerberg's vision... if he is being honest that is
Q: is Zuckerberg proposing we have our own personal AI machines just like we used to have our own PCs in the 90s ?
Yes, that is exactly the historical parallel Mark Zuckerberg is drawing.In his manifesto, The Future is for Everyone, Zuckerberg argues that computing power must be distributed directly to individuals rather than being locked up in a few giant institutions.Why the 1990s PC Analogy FitsLocal Control: Just like the 1990s shift from centralized corporate mainframes to personal computers (PCs), Meta is pushing for decentralized "personal superintelligence".Home Hardware: Meta's latest model, Muse Glimmer, is specifically engineered with 30 billion parameters so it can be downloaded and run locally on a home laptop using a single graphics card.Individual Ownership: Zuckerberg believes the future should not rely entirely on a few closed, centralized corporate clouds (like OpenAI or Google). He wants people to own and run their own private AI systems.Key Differences from the 90s PC EraContinuous Autonomy: Unlike a static 90s desktop that sits idle until you type on it, these personal AI systems are envisioned as 24/7 autonomous agents that constantly work on your behalf.Form Factor: Instead of a beige box on a desk, these personal machines will be integrated seamlessly into everyday variables like smart glasses to provide real-time, context-aware assistance.
Yes, that is exactly the historical parallel Mark Zuckerberg is drawing.In his manifesto, The Future is for Everyone, Zuckerberg argues that computing power must be distributed directly to individuals rather than being locked up in a few giant institutions.Why the 1990s PC Analogy FitsLocal Control: Just like the 1990s shift from centralized corporate mainframes to personal computers (PCs), Meta is pushing for decentralized "personal superintelligence".Home Hardware: Meta's latest model, Muse Glimmer, is specifically engineered with 30 billion parameters so it can be downloaded and run locally on a home laptop using a single graphics card.Individual Ownership: Zuckerberg believes the future should not rely entirely on a few closed, centralized corporate clouds (like OpenAI or Google). He wants people to own and run their own private AI systems.Key Differences from the 90s PC EraContinuous Autonomy: Unlike a static 90s desktop that sits idle until you type on it, these personal AI systems are envisioned as 24/7 autonomous agents that constantly work on your behalf.Form Factor: Instead of a beige box on a desk, these personal machines will be integrated seamlessly into everyday variables like smart glasses to provide real-time, context-aware assistance.
so, my next at home $3500 PC will not be a gaming machine... it'll be an AI machine.
Nah, he sounds out of touch and delusional. And frankly it's a toss-up between him and Musk who I'd trust less to lead the AI revolution. I hope meta and xAI both keep failing to make significant progress.
On August 19 2026 09:06 iPlaY.NettleS wrote: More info on the AI answering 911 calls in New Orleans via Google AI.
How the New Orleans 911 AI Works Limited Use Case: The system only handles repeat calls about vehicle accidents that are already reported.
Strict Activation Rules: The AI only answers if all human call-takers are busy and the call comes from within 200 meters of a known crash.
Immediate Transfer: If you are reporting a new emergency, or if you say anything other than "Yes" to the accident location prompt, the system routes you to a human dispatcher
.
It is an interesting thing to scale. One of the largest issues with these emergency response lines is prank calls or mistaken calls (such as phones calling by themselves when you are skiing). Answering and then transferring to a person if there is actually somebody talking at all could be another use case?
If they get reliability up high enough you could perhaps screen people using it incorrectly as well. Had a bus driver talk about a kid calling in an emergency when he was told that he had to pay for his own buss fare to get on...
Detecting whether there is someone speaking isn't AI, that's a bandpass filter, something that could've been done since roughly the 1930s. I'm guessing that this isn't used is because there's a risk of discarding a call where the user legitimately isn't talking because of something else, but will start talking later. But if they aren't already using something like that, using AI is a strange leap.
AI allows for flexibility, it interprets the intent and also can be multi linguistic.
Yes, there are obviously uses for LLMs in any callcenter. But specifically for detecting whether there is someone talking on the other end, it is extreme overkill (and probably even outright worse at the job).
Similarly, a lot of people say that AI is great for automation when they could've done the exact same thing 30 years ago with a cron job. The only thing the AI did is provide a user interface around the geeky cron syntax. Google also has a video of how wonderful AI integration is in Gmail, and it demonstrates it with a task to send the user a push notification when they get an email from their boss. Once again, lots of capacity to use AI in email. The second example in the video of summarizing stuff, for instance, is clearly something that could not be done well before we had AI (not only LLMs, but the basics of agentic AI here: to summarise well for a user, you need to know what the user thinks is important).
On a totally unrelated topic, I read this: https://www.economist.com/by-invitation/2026/08/20/humanity-has-the-debate-about-ai-consciousness-backwards To me it feels self-evident, in the same way Douglas Hofstadter's first presentation of the idea of conscience as a "strange loop" (if you prefer Dennett, the ideas are fully compatible, as they wrote Gödel, Escher, Bach together, but I'd attribute the phrase "strange loop" to Hofstadter). However, unlike Blaise Agüera y Arcas, I don't really see much evidence of such strange loops occurring in LLMs. If we include the agentic harness then maybe, but I feel like it's still far too simplistic. I can see how just adding more compute, more storage and more capabilities to act and interact with the environment could lead to emergence of "consciousness", though. Or at least to something that we'd find functionally indistinguishable. And then we get into Chinese Room arguments.
I kind of agree but also disagree at the same time. It's an overkill, but it's also an infrastructure for more features.
For email AI integration, it's been very useful for sure. It has kind of transformed how I interact with my email inbox now.
We have like 6 key platforms: artwork storage, airtable for task, emails for conversation (and external job progress), ERP system, google sheet, excel etc. As long as some of these can send an email, like my google sheet tells me which shipment is landing to port within 3 days, I can just ask outlook to give me an executive report, prioritizing actionable emails.
- Apple just announced the new mac mini and they are possibly the best value local AI machine right now. The m5 MAX chip unbinned has 600+ GB/s bandwidth, value proposal wise, it beat out any other AI mini box EASILY.
It gets close to RTX 6000 Ada bandwidth, around $7K and you are getting just 48GB on that. a 256GB M5 MAX, with 1TB storage is like $5K. (Education price)
The m5 Ultra chip is even more insane, bandwidth 1.2TB/s, gets very close to blackwell. Around $9k for a 256GB model (Education price). Blackwell is like $9K.
On August 25 2026 22:26 JimmyJRaynor wrote: i like Zuckerberg's vision... if he is being honest that is
Q: is Zuckerberg proposing we have our own personal AI machines just like we used to have our own PCs in the 90s ?
Yes, that is exactly the historical parallel Mark Zuckerberg is drawing.In his manifesto, The Future is for Everyone, Zuckerberg argues that computing power must be distributed directly to individuals rather than being locked up in a few giant institutions.Why the 1990s PC Analogy FitsLocal Control: Just like the 1990s shift from centralized corporate mainframes to personal computers (PCs), Meta is pushing for decentralized "personal superintelligence".Home Hardware: Meta's latest model, Muse Glimmer, is specifically engineered with 30 billion parameters so it can be downloaded and run locally on a home laptop using a single graphics card.Individual Ownership: Zuckerberg believes the future should not rely entirely on a few closed, centralized corporate clouds (like OpenAI or Google). He wants people to own and run their own private AI systems.Key Differences from the 90s PC EraContinuous Autonomy: Unlike a static 90s desktop that sits idle until you type on it, these personal AI systems are envisioned as 24/7 autonomous agents that constantly work on your behalf.Form Factor: Instead of a beige box on a desk, these personal machines will be integrated seamlessly into everyday variables like smart glasses to provide real-time, context-aware assistance.
so, my next at home $3500 PC will not be a gaming machine... it'll be an AI machine.
Nah, he sounds out of touch and delusional. And frankly it's a toss-up between him and Musk who I'd trust less to lead the AI revolution. I hope meta and xAI both keep failing to make significant progress.
no single person will lead the "AI revolution". the revolution is already happening. it has changed the lives of countless coders already.
One of my employee's confessed to me he is afraid to spend any time learning a new programming language lest an LLM render his learning useless. Fascinating times! I am going to have to carefully prepare a rebuttal for him. Before I do though I have to spitball a few things and think a few things through.
Right now, I'd say coders are in the same position mathematicians were in when tools like Maple arrived on the scene in the mid 1980s. Back then, and, now today a whopping 40 years later it remains vital to learn the foundations of mathematics in excruciating detail. In 1985, Maple replaced everything a first year calculus student could do. Its still important to be able to understand the Fundamental Theorem of Calculus in depth and detail. It is still important to understand Algebra.. what better way to do that ... than to do the same math problems that challenged people in 19th century.
It remains vital to learn a programming language in great detail and depth in much the same way math students learn the minute details of calculus, algebra, combinatorics in 2026. Maple could do a large portion of the tasks associated with the math theories in 1985.
I suggest computer scientists continue forward in their learning journey in much the same way mathematicians do today. Coders need to start treating programming languages like disposable tools. That said, detailed mastery of one programming language at the start is mandatory. During the detailed debugging process . ..You cannot fix what you do not deeply understand.
To those afraid to learn a new programming language i borrow a passage from Aldous Huxley's Brave New World... "The experiments were abandoned. No further attempt was made to teach children the length of the Nile in their sleep. Quite rightly. You can't learn a science unless you know what it's all about."
I need to spend more time thinking about this issue. What a time to be alive. "Oh Brave New World.. with such people in it".
On August 26 2026 20:27 JimmyJRaynor wrote: One of my employee's confessed to me he is afraid to spend any time learning a new programming language lest an LLM render his learning useless. Fascinating times! I am going to have to carefully prepare a rebuttal for him. Before I do though I have to spitball a few things and think a few things through.
Right now, I'd say coders are in the same position mathematicians were in when tools like Maple arrived on the scene in the mid 1980s. Back then, and, now today a whopping 40 years later it remains vital to learn the foundations of mathematics in excruciating detail. In 1985, Maple replaced everything a first year calculus student could do. Its still important to be able to understand the Fundamental Theorem of Calculus in depth and detail. It is still important to understand Algebra.. what better way to do that ... than to do the same math problems that challenged people in 19th century.
It remains vital to learn a programming language in great detail and depth in much the same way math students learn the minute details of calculus, algebra, combinatorics in 2026. Maple could do a large portion of the tasks associated with the math theories in 1985.
I suggest computer scientists continue forward in their learning journey in much the same way mathematicians do today. Coders need to start treating programming languages like disposable tools. That said, detailed mastery of one programming language at the start is mandatory. During the detailed debugging process . ..You cannot fix what you do not deeply understand.
To those afraid to learn a new programming language i borrow a passage from Aldous Huxley's Brave New World... "The experiments were abandoned. No further attempt was made to teach children the length of the Nile in their sleep. Quite rightly. You can't learn a science unless you know what it's all about."
I need to spend more time thinking about this issue. What a time to be alive. "Oh Brave New World.. with such people in it".
this is such a naive comparison. to do higher mathematics you need linear algebra and real analysis. and no, learning how to multiply 2 matrices, how to find eigenvalues or how to take a derivative has nothing to do with higher mathematics. maple/mathematica and other tools are mostly used for numerics, which are a very special case of "higher mathematics" and not the norm. higher mathematics is all about proofs, forget about numbers and arithmetic.
I don't know how is higher computer science but I doubt it has much to do with knowing the limitation of one programming language, and more to do with logic and higher mathematics.
then again, if you read Peter Sarnak on AI's Impact on Mathematics, you will see that from a math parspective, best ai models, could only brute force or use methods alredy known to break conjectures. they never invented anything new. Moreover, there is also the human factor. if you read lastest Terrence Tao presentation (the slides) you will see that he proposed, no proof will ever be accepted if the human is not a world expert in that branch, so there goes all your wet dreams about how ai will revolutionize anything.
Spot on regarding the debugging aspect. LLMs are great for churning out boilerplate code fast, but the moment you hit a edge-case memory leak or race condition, relying on AI without knowing underlying memory management or concurrency models gets you nowhere. AI speeds up implementation, but architecture and debugging still require real domain expertise.
Not into software engineering, but plenty of predominant software engineers are saying the human value will be at judging the work of AI more than anything else. And we are already seeing more talking about less prompt, and let AI run itself.
Also semi related to AI, the DLSS5 potential is absolutely insane.
It's hard to believe how quickly it can "generate" some sort of AI filter and have better texture and lighting etc.
Just want to share an anecdote of my recent AI use: I discovered astrophotography is fun. And most of that is thanks to Claude. I have enjoyed taking photos since I was a kid, and I've been getting better equipment as I get older, but I have never really put a whole lot of effort into developing the pictures. If I couldn't do it in a few minutes in Lightroom, then the photo was probably not good enough to bother with. Only very seldom did I actually really sit down and try to adjust the tones separately, put masks to differentiate bits of the picture: mostly I just relied on the fact that with digital I had shot 3 or 4 different versions of the same thing, and another photo was already starting from a better point anyway.
But I live in Spain, and I hiked up a mountain to see, and photograph, the total solar eclipse. Now I knew I needed a filter to shoot the partial eclipse, but I didn't really know much else, so I asked Claude what I needed to focus on for eclipse photography and it told me I should try to do a burst right at the start, and the end, of totality, to try to capture Baily's beads. And it told me I needed some good brackets of the corona. The former, I sort of managed, not as well as I'd like, but the latter was pretty easy, and when people asked me for pictures of the eclipse I pulled the 1/60 photo of the corona that looked pretty decent without any processing (because I was on holiday) and sent it, as well as a cool sunset with a bite out of it. But then I got home from the mountains and figured I should develop the pictures, but didn't have the faintest idea how to combine the bracket into a decent HDR in Lightroom, so logically, I asked Claude (not claude code, just claude.ai). It first told me to do what I had already tried, and lightroom had told me it couldn't do, which was to just merge them all in Lightroom. But because the photos all differ, and the only thing that matches is the lunar disk, which isn't enough for lightroom to lock onto and doing it manually is a pain in the ass, claude.ai was soon writing a script to do it in python. I'm fluent in python so could follow along, and saw it was basically importing the raw images to numpy and then doing a bunch of math (most of which I didn't follow completely but could get the gist of) to find the moon, translate the images a few pixels back and forth to match, and then merge them using Debevec-Malik radiance averaging, which I had never heard of. Anyway, from the absolutely washed out 1/2 and 1s images down to the 1/500 absolutely faint and useless image of the inner corona, it (we?) created an absolutely gorgeous rendering of the chromosphere and K-corona that blew my mind and I didn't know was hidden in the photos.
And that is when the fun started. Because now I realized that what we were doing was just signal processing and while signal processing is not my main field, it's adjacent (I'm a data scientist) and I know enough to grasp the basics and dive in: at the end of the day, most common filters are pretty simple concepts, but didn't have a clue it could be applied to my amateur photographs to tease out as much structure as I'm finding. And there's no way I would have even managed to get started without AI. I would have taken my 1/60 photo, polished it up, and thought that was about the best I could do with my amateur setup.
And not only am I going quite deep down the rabbit hole on signal processing, I am also learning a lot more about photography and the sun as I go. Sure, I'm not writing any of the code and I'd probably learn more about all the specifics, like how demosaicing works or what the chromosphere is, if I did a course or read a book. But this is more fun
If someone wants to see, DM me, but I don't want to share it publicly, because I'm considering submitting it to a competition. Honestly, I like my rendering quite a lot more than this one on APOD (https://apod.nasa.gov/apod/ap260817.html). Of course, I didn't have it processed 4 days after the eclipse either, and timeliness counts too!
Anyway, here's the 1/60 photo alone, developed the way I would have if I didn't have Claude to teach me stuff
EPT
![[image loading]](https://i.imgur.com/XwXibVn.jpeg)
