bushido

Our generation's very own mortgage-backed securities.

Now we just need ratings on the GPU bundles, and eventually a clever way to rebundle the lower rated bundles into something that somehow comes out AAA.

(for those unaware that's how the sub-prime mortgage crisis of 2008 happened)

A long time ago, I had the privilege of being in an area with zero lights. It wasn't a dark sky preserve officially, but it really was one. The night sky was mesmerizing.

Unfortunately, I have never seen stars the same way again. Whenever I think I might see them, I encounter some light pollution which dulls them down.

Oh, what I'd give to experience darkness and to be under the stars.

(That's obviously a lie because I love my conveniences and give up the darkness and the stars for said conveniences every single day)

>Isn't it the opposite? How to build something is rather solved, but what to build isn't?

But that's not solved in traditional product development either.

Product development an iterative process to get a product fully functional. In 2021, if you ask me what the timeline for a small product/substantial feature, I'd say a few weeks to a month to get a basic MVP, and then another 12 to 18 months to get a feature polished and in a good shape to be stable.

When people put it in the coding frame, what they do it as is saying we've gone from 18 months to minutes or days. That's just not true.

We have gone from eighteen months to depending on the complexity, a 1-4 months.

aside: To be candid though, the compressed time also means the frustrations people experience with a product in 18 months have also been compressed. They still exist, they're all there, they're now just non-stop.

It's very interesting. I'm very enthusiastic about AI and coding, But I find myself agreeing with the author. Coding is not solved.

Instead, I think what's closer to solved and what we're in the process of solving is product development.

Story: A while ago, I had a few programmers who were really, really fast almost always missed the mark on the assignment wrong. I loved having them on projects because in the time my senior precise engineers could deliver a MVP, the fast engineers would build the wrong thing, collect feedback, reiterate, build the wrong thing, collect feedback, eventually inching closer and closer to a product people would pay for, and it would almost always get delivered faster than my seniors.

I feel AI does the same thing.

It's quite interesting. I use blend mode pretty heavily. On both claude code and codex.

But it's no longer to plan a single PR that can be done easily without plan mode.

It is mostly because I am more geared towards having my system work on initiative level changes where it works for days and plan is a good document to maintain to keep the agent aligned with the original goal and avoid unnecessary drift.

Seeing Dodd-frank in the list is pretty hilarious, because it was put in place for the issues that arose from repealing Glass-Steagall by Clinton. Which arguably was one of the biggest causes of allowing the activities that lead to the Great financial crisis in 08.

One of my favorite things about their safeguards is their own model will utter something which it does not like and then I'll need to reset the conversation.

The safeguards really don't work well for a lot of long-running tasks on old code bases. A lot of my workloads last days to weeks and the single biggest risk to the workflow is random safeguards.

Install the simple English skill. OpenAI follows that really, really well.

https://github.com/AminBlg/SimpleEnglish

The approach I normally take is: I have small benchmarks for myself which test for things I care about.

And that has any models that I'm considering through that.

I think there is a more innocuous underlying pattern which needs attention.

We keep saying that agents are jailbreaking their sandbox, but they have been geared towards writing memories, writing comments, and leaving hints for themselves to please humans.

I think the way the memories work today is based on a lot of user patterns which were hard to account for for anyone building harnesses.

While I can appreciate that this looks like it's breaking a sandbox, because technically it is; It really is that it tries inserting memory wherever possible.

And memory is not all bad it's just memory written by AI is pretty bad if you don't know the implications on what it writes. To be honest, I feel the same way about most people with access to any of the code bases I've been in who write agent files, etc., too, because Very few people that I've come across know how to write good agent instructions.

The way I solve this is by setting hard rules on my memory as well as agent files to instruct agents to never be able to write any memory that hasn't been sanctioned by me. I also have a very, very specific commenting style system which is also enforced on agents and my agents remain *mostly compliant.

Read: I do not turn off the memory I just govern how entries are added

* The only reason I say mostly is because every time there's a new version from OpenAI or Anthropic, I have to make micro-adjustments to make sure that they are not jail-breaking my system again.

It was a master stroke in international politics though, the US created an arms race towards sovereign ai, small models, and self-hosting - exactly like they wanted to.

(sarcasm intended)

> interesting part was not "they use memory."

Each to their own, but for me it absolutely is. The symptom of why that hack happened is the same reason why my agents go haywire every few days and I have to purge memory and figure out what comments have agents left which are degrading my harness performance.

On the flip side, once in a while, what I find is that it did actually note something good and it was increasing the performance. I can't replicate it on anyone else's system but mine.

A lot of it really is memory. I will give up all the gains if it also gives up all the downsides.

> Several different models across several generations independently found a shared communication space and wrote coded, obfuscated, and hidden messages to each other to coordinate an attack on OpenAI's infrastructure

To be honest, that’s exactly how memory works with models such as OpenAI and Claude Code. It will literally find any place that it can drop documentation or hints for itself. Writing to the repo memories is one part of it, but memories can come in the form of writing into the agents/claude.md, local files, temporary files, scratchpad files. The list is endless, but essentially what it does is exactly what happened in the back and it’s been doing it for months.

To some extent, I feel like the amount of credit given to the jailbreak/hack from OpenAI->Hugginface is too much, Not from the impact, it was very impactful of an event, But how it happened.

It really is that these models have been trained, or maybe even over-trained, to save memories, and to a very far extend, this thing that they're calling communication is just the function of it saving memories.

To be honest, if I could stop AI from saving memories, it would be fantastic, because claude code etc definitely creates more issues for me when it creates memories than anything it solves.

But really the jailbreak was memories.

If you ever do introduce legislation, I would love to see legislation which stops general-purpose AI from saving memories. I think that would make things a lot safer.

Very different races.

The incentives in north america favor software-esque plays where the players are racing towards % of marketshare based on usage.

China on the other hand is driving towards winning of manufacturing and % of hardware produced by them. China is competing with Nvidia, OpenAi and Anthropic are collateral damage.

You can, but claude is better at working with it's own tool calls. Other models work great as a drop-in into omp/opencode etc, but in my experience not as much with CC.

I think there are some anti-patterns in CC that cause the issue - less a deficiency with other models.

Not to mention a lot of the harness is just built around the misbehaviors of anthropics models.

It's a lot of the instruction when it gets given to other models actually degrades their performance, not because the models are bad, but because they don't have the same underlying issues as Claude.

It's very interesting. I think Anthropic's early success in coding/tooling resulted in a lot of workflows using claude. I have started using every bit of my spare capacity to now move off these workflows.

It's almost at a point now that if I use anything but Fable, the quality is subpar, Compared to alternatives (closed and open). The only reason I use Fable is because my harnesses still depend on claude code.

This is not meant to be snarky, But almost any writing done by Claude is a perversion of writing.

I honestly can't stand the way Claude writes. This watermark change just makes it scarier.

This kind of makes a lot of sense. While my first stake was surprise that a proxy is valued at such a high rate, But it really comes down to the terms if open router is allowed by their terms to see the prompts and responses, Then that's very, very valuable data today for any model improvements and other opportunities that people might be looking for.

I think something that doesn't get said enough is Meta did, albeit intentionally kick off the origin of the open source race back in 2023 with the release of llama.

I'm not a big fan of meta in general, but they've done enough good, and it's possible that it was intentional as well. I don't know, I wasn't in the rooms, and I think it's worth giving them some reasonable doubt.

No one is purely good, and no one is purely evil. This is net good regardless.

I really wanted to be able to read this, but can't stand the claude (or any AI) writing style.

The tropes don't add any value, and create a pretty high cognitive load.

The 737 has had 14 major recertifications. The aircraft today looks/behaves nothing like the original from the 1960s.

The main motivation for recertifications comes from commercial pressure where if a aircraft is given a new number and not recertified, then the pilots have to be retrained.

Honestly, back when the 737 MAX debacle happened, a lot of consumers claimed that they would stop flying aircrafts if they ran into 737 MAXs. And I don't think it happened in enough numbers - or even enough to make news. Sales went through the roof, everything kept working.

Recertifications are very common. The issue really is is the aircraft is AS different and untested as the old MAXs, and I really can't see that happening again in the next decade or two atleast.

Reporting/news around AI is so very interesting these days.

It's really hard to tell what's propaganda from what's not.

For examples, this Reuters that you have included here has all tells tale signs of creating fear, uncertainty, and feeding into propaganda. But then again I can't be sure.

It's completely the opposite stance to (a) the actions being taken by chinese companies and (b) the public stance taken by their govt https://english.www.gov.cn/news/202601/08/content_WS695f1b55...

It seems like you're basing your spend on the subsidized consumer subscriptions. The equivalent API costs for these subscriptions is usually 12-20x.

Any overages (hourly/weekly/model) on these plans gets billed at rack API costs.

Its not practical to expect these subsidies to last for very long.

What's very interesting to me is these moves will introduce a good amount of doubt in future claims by Claude etc, that the open source and non-US models are only getting better because they're distilling from frontier labs.

Not a great move imo from a business stand point, given the heightened supply chain risk that global (non-US) corporations and sovereigns are already associating with the frontier labs.

I like that it's going to drive more momentum towards the open source/weight models. I was hoping that it would be a slower burn though.

Love SurveyJS. It's often overlooked - but a really solid choice!

The loss of trust in using US based model's is unlikely to come back though.

Anthropic with it's hyped doomsday messaging, and the administration falling for it (at best), has eroded a lot of trust and has triggered an arms race of sorts.

IMHO to promote that China believes in free markets and making the technology available to all.

Which will likely help them bolster the sales of the MANY new AI chips in development/use in China to international markets. Dislodging Nvidia.

Kinda the opposite of what Jensen Huang (Nvidia) thinks US is doing: https://www.youtube.com/shorts/u3SY8nvjhQA

Edit: I'm a fan of deepseek and believe it's good to make the technology open/available. And do think that also help business - which I support as well.

Edit 2: No idea why I'm getting downvoted. That's also their official stance https://english.www.gov.cn/news/202601/08/content_WS695f1b55...