web dev, digital marketer, content creator, graphic designer, x1 solo-founder, x2 co-founder | dad/husband | prev. nightclub bartender, property manager
NW FL
Joined November 2015
- Tweets65.4K
- Following887
- Followers2.5K
- Likes11K
Pinned Tweet
I regret calling this a "memory" tool. It isn't one.
A memory is the output of processing information, not the information itself. To become a memory, information has to be consolidated and summarized into something new that stands in for the original.
That step is where things go wrong. Memories are easy to poison, and a poisoned memory infects everything it touches with wrong context. You can't check it either, because the source it came from is gone.
agent-history doesn't do that. It makes no assumptions about your information and creates no memories. It doesn't need to.
It indexes your past agent conversations and lets your agents search them by meaning. Results are ranked by a blend of semantic similarity and keyword match, with ties going to the newer conversation, so a later correction beats the thing it corrected.
Embedding, search and ranking all run on your machine. The only provider tokens you spend are when your agent reads the results.
A memory gives you one fixed summary. This lets you go into the source conversations and find exactly the context you need, in your own words, as it was actually said.
Just ask the way you'd ask a person. You don't need to remember keywords or which project something happened in, because it indexes agent conversations from across your machine by default: Claude Code, Codex and OpenCode.
Tool calls, command output and other terminal noise are stripped out before indexing. That makes it easier for you to read and cheaper for your agent to process.
this guy is playing 6-7d chess
Dario Amodei: some tokens Claude produces are worth a few cents, others could be worth tens of millions of dollars.
His examples:
1. Telling someone to restart their Mac: the model has said it 10 million times, worth a dollar or less
2. Telling a pharma company to move an aromatic ring to the other end of a molecule: potentially tens of millions
3. Expect pay-for-results pricing, or labor-like pricing by the hour
4. It's a new industry, so many models will be tried
Today pricing is simpler: per the S-1, usage-based pricing was 80% of H1 2026 revenue and subscriptions 20%.
Full breakdown here:
theaiopportunities.com/p/ant…
I see a lot of people saying that claude duet is unnecessary because you can just use claude code or claude app with projects, and I don't disagree.
The real value, I think, comes from the actual UI. It's actually built for exactly this kind of interactivity.
One of my favorites is how it displays email drafts. It's just so clean with a send it or edit it option.
Emails is one of the things I've hated the most with agents up til now, out of the box of course. Obviously you can build your own ui, but discounting that agents have sucked at actually showing you what an email will look like, with the subject and recipients neatly and in a way that mimics what it will look like when it's actually being sent.
I have spent more time fine-tuning how agents handle composing emails than I have on any other specific thing, I feel like.
So yea, just my thoughts.
honestly, this is exciting because it's forcing us to hunker down our environments and ensure that we're inspecting traffic in/out on a regular basis, instead of just when we're trying to look for something specific or audit.
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
I really do love claude duet, but I think the trick is understanding when to use it vs when to dive in to a claude project in the app or use claude code.
Luckily, I built my agent history plugin which gives my duet assistant access to all my conversations from my local and cloud harnesses via mcp, so I can treat Duet like an actual assistant, rather than try to brute force everything through a single channel.
d3layd retweeted
You can see in this trailer Nathan doing something weird with his tongue and his tooth. In an interview, Nathan explained that when he's trying not to laugh or break character, he presses his tongue against his molar really hard. So him pressing his tongue against his molar really hard reveals that Nathan knows that Elizabeth Holmes is a lying manipulator, and it also indicates that Nathan is also trying to manipulate her and is in character. So then the way Nathan acts in the documentary is an act, he's conning a con artist.
d3layd retweeted
nous girl can do more than just research??
yeah, she's funny now. just tell your agent "hook up memelord mcp"
abliterated deepseek is a game changer.
pennies on the dollar to actually give me actionable insights that I can hand off to my agents to implement without having to worry about tripping safety classifiers.
Claude Duet has been a massive game changer for me.
I've always been ahead of the curve when it comes to working with agents, and Claude Duet is one of those things that makes me really happy about that.
Summer of 2025 I was having multiple agents coordinate over a shared markdown doc locally.
Early 2026 I was building custom integrations for OpenClaw. I was a big fan of building out channels in discord/slack that allow for specialized agents to live.
By mid 2026 I was building out structured systems for my OpenClaw and Hermes agents to use to maximize determinism and predictability, while reducing unnecessary token spend.
So when Duet rolled out, it only took a few minutes to connect it to all of my existing infrastructure and tools and has completely changed how I interact with my agents and Claude as a whole.
It's configured to reach out and interact with my Hermes agents, it uses a single source of truth, and it's great at actually managing all of this context.
It's finally come around full circle to my original vision of having a single orchestrator agent that can invoke and coordinate between other agents to maximize productivity and reduce wasted effort.
IDK, I'm just excited.
Thanks for having my @AnthropicAI
Just want to shout out my boy @skeptrune for hooking me up with that abliterated deepseek from @morphllm for research purposes and internal red/blue teaming.
Absolute legend.
almost got a chatgpt pro plan but then talked myself out of it cause I would rather pay for another claude account if I needed to
fuck
Dario Amodei: some tokens Claude produces are worth a few cents, others could be worth tens of millions of dollars.
His examples:
1. Telling someone to restart their Mac: the model has said it 10 million times, worth a dollar or less
2. Telling a pharma company to move an aromatic ring to the other end of a molecule: potentially tens of millions
3. Expect pay-for-results pricing, or labor-like pricing by the hour
4. It's a new industry, so many models will be tried
Today pricing is simpler: per the S-1, usage-based pricing was 80% of H1 2026 revenue and subscriptions 20%.
Full breakdown here:
theaiopportunities.com/p/ant…
d3layd retweeted
OpenAI just casually announced that its AI has disproved the Exponential Time Hypothesis.
One of the biggest assumptions in theoretical computer science, closely related to P vs NP and used in hundreds of research papers.
OpenAI says the proof was formally verified in Lean.
And the craziest part? The researchers now examining it include Impagliazzo and Paturi, the guys who originally proposed ETH.
actually insane
