Ray Villalobos ✝️ retweeted
security architecture fundamentals are more important than ever before
if you’re working for a company in cybersecurity, you should be evaluating your options to abstract your companies services from the internet
what i mean is: why allow an unauthenticated anonymous agent of an attacker to interact with your systems?
consider your architectural levers now, you’re going to need them
186
149
1,920
135,443
Last month Linus Torvalds pushed a kernel branch named noref-filemap-read-with-llm and credited Claude Code in the cover letter. What happened next is the part developers should steal.
He went after a real Linux page cache bottleneck Matthew Wilcox raised
go.raybo.org/129BW
He went after a real Linux page cache bottleneck Matthew Wilcox raised
go.raybo.org/129BW
0
0
1
190
We might find out that the type of worker we need for the next century is a completely different type of worker
1
0
2
189
The real danger of falling behind right now is competition from a new category of companies that don't have your company's red tape.
Most corporate environments have entire departments built to prevent new things from happening at any cost, sort of like the US Congress. Meanwhil
Most corporate environments have entire departments built to prevent new things from happening at any cost, sort of like the US Congress. Meanwhil
0
0
0
102
The amount of fun I have working with AI is ridiculous. Hope you're enjoying it too. Also the workflows I'm using are insane.
0
0
2
156
Loving these simple illustrations I'm able to do with GPT Image 2.6, but I'm noticing a few more small mistakes like the double nose on the dog. Definitely more than in Image 2 @ChatGPT
0
0
0
116
Faster is nice. Staying consistent is what makes an image tool useful, and that's why we got a nice GPT 2.6 image update.
OpenAI just shipped Images 2.6 in ChatGPT, Codex, and the API. It generates images about 50% faster, adds GIF support, and lets you save character references
OpenAI just shipped Images 2.6 in ChatGPT, Codex, and the API. It generates images about 50% faster, adds GIF support, and lets you save character references
2
0
2
224
Believe it or not, models can edit other models without access to their code and make big benchmark jumps, according to MIT research.
Researchers gave GPT-6 Astra the raw weights of Qwen3.5-0.8B, a fairly small model, and asked it to make the model better at math. There was no t
Researchers gave GPT-6 Astra the raw weights of Qwen3.5-0.8B, a fairly small model, and asked it to make the model better at math. There was no t
0
0
0
163
Revoking an agent's tokens one by one can lose a race to a refresh in flight. Review of the MCP server I built caught it. The fix, now live: write a revocation record first, then check it on every read and issue. Removing a connector in your AI app doesn't revoke it server-side.
5
0
2
205
A faster model won't fix a slow agent if the model was never the slow part.
Martin Alderson wrote up a year of making coding agents faster, and one 60-second task shows the problem:
Here's where the time went:
- 24s of inference: reading context and gen
go.raybo.org/1252q
Martin Alderson wrote up a year of making coding agents faster, and one 60-second task shows the problem:
Here's where the time went:
- 24s of inference: reading context and gen
go.raybo.org/1252q
1
0
0
109
ChatGPT is changing the way they charge for tokens, and if you're a coder, you're not going to like it.
Starting in December, on ChatGPT Plus, Pro, and Business plans, ChatGPT and Codex will draw from the same allowance. That didn't used to be the case, so my chats had a differe
Starting in December, on ChatGPT Plus, Pro, and Business plans, ChatGPT and Codex will draw from the same allowance. That didn't used to be the case, so my chats had a differe
0
0
0
170
A cheaper model can match a flagship SOTA model on reliability, but only if you give it a way to check its own work.
I can tell you from experience that a solid harness will always outperform a smarter model.
I can tell you from experience that a solid harness will always outperform a smarter model.
1
0
0
147
If I need a code review, design, motion graphic or 3D, I just tell Grok to fire up one of the CLIs like Codex or Claude Code and use the best tool for the task, then continue triaging as usual.
1
0
0
42
Why not the open source models? My own GruntBench shows they're good at looking good on intelligence benchmarks, but not great at getting real work done.
Forget the benchmarks. Pick a workhorse model. Then give it a great harness and watch it crush your work.
Forget the benchmarks. Pick a workhorse model. Then give it a great harness and watch it crush your work.
0
0
0
13
My programming harness is under constant evolution, and it lines up with some of the findings from the Claude Code team. The days of most human code reviews are done (::ducking::)
Where Claude writes about 90% of the code, merged PRs more than doubled from June to September, and
Where Claude writes about 90% of the code, merged PRs more than doubled from June to September, and
4
0
2
170
The model is already good enough for real work. But most companies still aren't using it that way. Here's why you need to learn the missing link.
The new job title getting popular is Forward Deployed Engineer.
The new job title getting popular is Forward Deployed Engineer.
1
0
0
118
Raw model intelligence is no longer the bottleneck. The friction is in the loop. In fact, Anthropic's study on internal agent use found that 80% of interactive session time was agents sitting idle, waiting on human review.
1
0
0
30