Jake

ad astra | e/acc

[Jake @FindingUrPasswd]

Honestly a rare political tweet here but massive s/o to Chicago + IL state reps on not caving to unnecessary demands from a family worth $8 billion. The Bears will do incredible brand damage to a historic franchise all over property taxes if they go through with their move

[Jake @FindingUrPasswd]

@thedawgspodcast

[Jake @FindingUrPasswd]

Big time 👀

[OpenAI @OpenAI]

OpenAI frontier models and Codex are now generally available on AWS, giving enterprises a new way to build on Amazon Bedrock with OpenAI through the security, compliance, and governance workflows they already use.

This is also the beginning of a broader expansion of OpenAI capabilities on AWS, including future availability for cybersecurity capabilities like Daybreak.

https://t.co/vMws0YU6Q3

[Jake @FindingUrPasswd]

@Browns Is making generationally bad decisions the only way we get attention now?

[Jake @FindingUrPasswd]

My TL was flooded with complaints about the Claude Code token consumption bug a month or so ago but I haven’t seen a single tweet about the Codex token consumption bug 😂

[Jake @FindingUrPasswd]

@karpathy

[Jake @FindingUrPasswd]

I spend a lot (and by a lot I mean a lot) of time doing development & imo these are 5 things you should be doing to effectively dev with AI:

test driven development + strict typing
the model always has a contract to code against. without it you're just vibing into the void and hoping the output is correct, a simple suite of tests generated by the models themselves go a long way in reducing small syntactic errors

plan driven development
write a plan file that breaks implementation into phases with explicit manual validation checkpoints at each step. pause, verify expected state, resolve deviations before moving forward. this keeps the model's context sharp and prevents compounding mistakes across a long task

strict code readability/simplicity instructions
there's an old NASA paper on this that's worth a read:
(https://t.co/H3HE02L5vJ)
short functions, bounded loops, no recursion, check every return value. models will write clean code if you tell them exactly what clean means to us

be adversarial to your own design
have the model poke holes in your decisions before you commit to them. models will make hacky assumptions if you let them & I've found this surfaces those early rather than debugging them later

version control discipline
feature branch per task, commit at each validated checkpoint. this gives you clean rollback points and makes it obvious where a deviation from the plan introduced a problem

[Jake @FindingUrPasswd]

Ironman is officially 6 weeks out ⏳😬

[Jake @FindingUrPasswd]

The viet dip from Phodega is a top 5 Chicago bite can’t lie

[Steven Heidel @stevenheidel]

SpaceX? the coding app?

[Jake @FindingUrPasswd]

The takes on your feed this week are probably running the full range, from "Mythos is the end of security" to "Claude Code is garbage and unusable now." Reality's almost certainly somewhere in the middle. Wrote up some of my thoughts on the matter here: https://t.co/0TFCej6ZLE

[Jake @FindingUrPasswd]

I got to write up our analysis on this one! If your org is using LiteLLM as an LLM proxy layer, this is a "rotate credentials now, ask questions later" situation

[NetSPI @NetSPI]

On March 24, 2026, a severe compromise was identified in LiteLLM, a widely adopted open-source LLM proxy framework. A compromise at this centralized layer exposes your entire LLM provider surface area. Read more: https://t.co/WnAjgouCsL

[NetSPI @NetSPI]

On March 24, 2026, a severe compromise was identified in LiteLLM, a widely adopted open-source LLM proxy framework. A compromise at this centralized layer exposes your entire LLM provider surface area. Read more: https://t.co/WnAjgouCsL

[Jake @FindingUrPasswd]

ycombinator taking an absolute PR beating this week with the Delve fiasco & now this

[Jake @FindingUrPasswd]

oh boy, if any of you are reading this and using liteLLM (or know your org is using liteLLM). Please monitor the GitHub issues thread (https://t.co/vWza9g5u9X) for remediation guidance. One of the few market viable options for this so the impact on this one is going to be large..

[Daniel Hnyk @hnykda]

LiteLLM HAS BEEN COMPROMISED, DO NOT UPDATE. We just discovered that LiteLLM pypi release 1.82.8. It has been compromised, it contains litellm_init.pth with base64 encoded instructions to send all the credentials it can find to remote server + self-replicate. link below

[Jake @FindingUrPasswd]

HUH

[艾略特 @elliotchen100]

论文来了。名字叫 MSA，Memory Sparse Attention。

一句话说清楚它是什么：
让大模型原生拥有超长记忆。不是外挂检索，不是暴力扩窗口，而是把「记忆」直接长进了注意力机制里，端到端训练。

过去的方案为什么不行？
RAG 的本质是「开卷考试」。模型自己不记东西，全靠现场翻笔记。翻得准不准要看检索质量，翻得快不快要看数据量。一旦信息分散在几十份文档里、需要跨文档推理，就抓瞎了。

线性注意力和 KV 缓存的本质是「压缩记忆」。记是记了，但越压越糊，长了就丢。

MSA 的思路完全不同：
→ 不压缩，不外挂，而是让模型学会「挑重点看」
核心是一种可扩展的稀疏注意力架构，复杂度是线性的。记忆量翻 10 倍，计算成本不会指数爆炸。
→ 模型知道「这段记忆来自哪、什么时候的」
用了一种叫 document-wise RoPE 的位置编码，让模型天然理解文档边界和时间顺序。
→ 碎片化的信息也能串起来推理
Memory Interleaving 机制，让模型能在散落各处的记忆片段之间做多跳推理。不是只找到一条相关记录，而是把线索串成链。

结果呢？
· 从 16K 扩到 1 亿 token，精度衰减不到 9%
· 4B 参数的 MSA 模型，在长上下文 benchmark 上打赢 235B 级别的顶级 RAG 系��
· 2 张 A800 就能跑 1 亿 token 推理。这不是实验室专属，这是创业公司买得起的成本。

说白了，以前的大模型是一个极度聪明但只有金鱼记忆的天才。MSA 想做的事情是，让它真正「记住」。

我们放 github 上了，算法的同学不容易，可以点颗星星支持一下。🌟👀🙏

https://t.co/Or3PVUv67q

[Jake @FindingUrPasswd]

@ITSecurityguard This is an absolutely excellent write up that more people need to read. Well done on this! 👏

[Jake @FindingUrPasswd]

@butleriano Atlanta in the top 3 instead of Chicago

[Jake @FindingUrPasswd]

Rip clawdbot lol

[Claude @claudeai]

New in Claude Code: Remote Control.

Kick off a task in your terminal and pick it up from your phone while you take a walk or join a meeting.

Claude keeps running on your machine, and you can control the session from the Claude app or https://t.co/er6Blrr63e