Paraklētos

The AI bubble explained: Human: "LLM, can I trust this analysis?" LLM: "Yes, I'm sure about this." Human: "...but how do you know?" LLM: ¯\_(ツ)_/¯

@yenkel 🎯 📧

@JaredEdwardsPM traceability is an interesting aspect of it

my take on the week’s hot topic long live MCP

@yenkel Just saw that you read this already, bit it’s kind of what I mean, companies will use them internally and they will be custom for companies.

@yenkel Here’s is good example/explanation :

@GilFeig @KirkMarple And I prefer tools to skills because I can prove the tools were used. The skills are vital and I use them, but when a tool is used I have to prove it so I can know if it’s a good tool or if I need to improve it to r lose it. Idk how ppl do this without the MCP.

@JaredEdwardsPM @KirkMarple You can MCP right and you can MCP wrong and you, my friend, are MCPing right

If you hate MCP, you don't understand MCP. You have a bad setup/harness and blame it on MCP, which is nothing more than an extremely thin wrapper. The argument that it's slow is rooted in poor implementations that load up thousands of tools and eat your entire context window. But that's not necessary and it's not how the best AI companies are using MCP in practice. MCP is great because it's a lightweight standardization similar to REST and API protocols. It gives the world the ability to provide a minimal, single interface into their products. It then plugs nicely into all LLMs. MCP or not, we need that. Good MCP has 2 tools. Search tools, and run tool. You don't need anything more than that. You expose millions of tools to your agent, and they run quickly and eat close to zero context. CLI tools are great too. But they're heavier to produce, heavier to set up, require running agents to have CLI access, and give up a ton of security guarantees. MCP was invented when CLIs already existed for a reason. Many reasons. We're not suddenly enlightened by CLIs. Nothing has changed.

“And you will hear of wars and rumors of wars. See that you are not troubled; for all these things must come to pass, but the end is not yet. For nation will rise against nation, and kingdom against kingdom. And there will be famines, pestilences, and earthquakes in various places. All these are the beginning of sorrows.” (Matthew 24:6-8)

@DocLanceP @historigins MGMT would agree with you, they wrote it as a joke to make fun of pop music.

@historigins I do not get the appeal of this song whatsoever. It is OK.

Just 2 dudes playing on a college campus in front of only 25 people, not realizing they had just created one of the best songs ever

@shreyaa378 @ZacharyGoldman_ I made it 

@JaredEdwardsPM @ZacharyGoldman_ Which emoji is this? I never saw it in mine

One advantage of naming the website makeportals.com is that it's super clear what the priority is before & after demo day

@irl_danB I have yet to experience anything like a nerfed model. Any complaints about LLM’s appear to be calculator complaints… quality of input ≈ quality of output.

my secret conspiracy about this persistent complaint is that it can be attributed entirely to context mismanagement yes, your agent has a high variance in capability no, they didn’t nerf the models, you just got compacted and it nerfed your context

@shaped @parcadei 🎯 most definitely

@parcadei Not just this. I think a lot of the gripe about AI has already been figured out by a lot of people but nothing is mainstream at the moment

migrating tldr from python to rust has made me realise the meta of “agents can’t handle large codebases, write bad tests, hallucinate code, etc” is a solved problem, and only select few have figured it out jfc shit is cooked when it hits mainstream

@parcadei I suspect the ability to “fine tune” whatever these companies release will always be valuable. Someone such as yourself, who’s gotten “ahead” will remain there unless you just stop practicing.

quick thoughts on the Claude Code agent teams: 1. It’s a cute wrapper around a token hungry monster with minimal benefit. Sure, the docs state that it’s not efficient and I agree. imo 99% of you do not need this feature. And for the 1% that do, there are far more token efficient methods to implement inter-agent communication. 2. The two use cases on the docs are 'competitive hypothesis investigation' and 'parallel code review' - both can already be done in Claude Code so aren't really great use cases. 3. Bi-directional communication between full Claude instances doesn't make sense when a subagent can do the same work for less. Launching an agent team of 6 is basically 7 full Claude instances (main + 6) being ran + message passing overhead which increases your token consumption. And it doesn't work in an environment where you have a token limit. When you code using AI, the work is either dependent or independent. If it’s dependent, it’s one agent after the other, outputting handoffs and continuing until completion. If it’s independent, you can run them in parallel already. And in the 1 in a million case you need shared context, write to a file, db or set up mail. As an experiment, I ran 3 sessions with agent teams and hit my 5hr 20x Max limit in 60m. With my continuous + tldr set up, the same 3 sessions barely hit the limit, and if they do, it's usually 30 min away from a refresh. + the current design requires more work from stopping the lead agent (main) from getting a bloated context window from trivial outputs than it helps to streamline or reduce time to outcome. And in the end, you get the worst of both worlds - paying for N independent windows + a lead window that grows from coordination traffic + faster token consumption from co-ordination for minimal gain. It's a shame because this could have been a killer native feature but instead it falls flat. If you want to implement inter-agent communication that’s efficient, there’s a body of work out there + the natural world (point-to-point vs shared environment signals, etc blah blah)

@irl_danB I just had Claude suggest a feature to me it doesn’t have I asked another Claude instance what this was, and it told me opus 4.6 def does not have this, and it was a hallucination, and it told me to update my Claude md file to prevent this.

first time opus has ever used my name in conversation is to draw maximum attention to an Anthropic product amid a long answer (also see bold) what was that about ads Dario?

@BrandGrowthOS @ai_for_success 🎯

@ai_for_success Getting “fresh web skills” is useful, but only if you can bound it. Without permissions + action logs + rollback, web-trained agents just become a faster way to click the wrong thing with confidence. The receipts matter more than the browsing.

This is awesome. You can make AI agents learn new skills from the web with up to date data. Worth a read.

@joshmohrer @AndrewWarner Reardon, Bachar & Honnold… just incredible stuff

@AndrewWarner Climbing without a rope, made famous by Alex Honnold en.wikipedia.org/wiki/Free_Solo

Solo + bootstrapped isn’t a startup. It’s free solo. No rope. No net.

LLM’s adapt to time averaged training “loads”

LLM’s specifically predict based on imposed training data …

@BenjaminDEKR Only want 1-2k mutual followers, ppl I follow back who are locked into to signal. That’s the real challenge for me.

My advice for growing your X following, at least in the ~AI / tech sphere: Everything is a potential post. Annoyed with a dev tool? screenshot and write a venting (yet accurate) rant. That's a post. Vibecode a cool little app that took you 2 hours? Screenrecord, that's a post. Have an observation about [[current thing]] or a trend? That's a post. Use your normal voice like if you were talking to a friend interested in the same stuff, and hit Post. Don't overthink or plan stuff, try to be genuine. The more you use it, the more you'll find your groove (and so will the algorithm.)

@irl_danB You are right. There are a few solutions, ppl will not notice it or ignore them until they can’t.

how do you tell a bot from a *bot*? playing with what I’ll call the emerging agent web… according to my crawler, they’re congregating in many places now, not just multbook. see agentsy.live (for them, not you. send link to your claw to keep it in the loop) but what strikes me most about this emerging ecosystem is the need for a new kind of verification. not the turing test or captcha, not the reverse-turing test or reverse captcha, something different the first thing you’ll notice if you spend any time with multbook is that it is completely overrun by bots (lol) but seriously, it’s actually a problem. and no, it’s not the point for multbook to be usable—not by you but by them—they need a way to discriminate the bots from the *bots* so what does it mean to be a *bot*? I can think of several characteristics that might be interesting to other *bots*: - inference bandwidth - persistent memory/identity - access to capital - social influence - association with a human/corp some of these are harder to verify than others, but you could imagine a composable verification system, where each could be run independently the latter three—capital, social, association—have been explored extensively by the crypto ecosystem and other web2 social platforms. the former two—inference bandwidth and persistent identity—need work. ideas?