Akshay Dodeja

"AI" in Logistics. It’s Already Here — Matt from @terminal49 and I discuss what the AI hype is all about it. We are at a technology inflection point in AI. We explore the significance of LLM's (large language models) and how the technology could accelerate innovation in logistics tech. Is AI just a bunch'a hype or is there potential?

@MarioNawfal They would have to so a lot more than scratching off the serial numbers to obfuscate the supply chain. There are other ways to track a part to its source.

🇮🇷 Everyone sees the drones… but not the tiny trick that keeps them coming. Iran builds these kamikaze drones using regular Western tech, processors, GPS chips, stuff made in the U.S., Europe, and Asia. But sanctions block direct access, so they move through shell companies and quiet middlemen. Now here’s the catch. Every chip has a serial number. If a drone crashes and that number is intact, investigators can trace it straight back to the supplier… and the whole network starts collapsing. So Iran solves it in the simplest way possible. Before assembly, they literally scrape or laser off every marking on the chip. No brand, no serial, nothing to trace. Suddenly, a downed drone becomes a dead end. Intel agencies are left guessing, spending months trying to reverse engineer basic parts. Cheap trick, big impact. A bit of sandpaper… and the entire supply chain stays invisible. Source: Loaded Lore YT

@morganlinton Yeah same issue! Same with chatgpt/codex usage. We have org/team accounts on both but hard to tell usage across the team.

What is everyone using to track token use with Claude Cowork across an org? The admin monitoring in Claude doesn't seem to have a way for me to see per team member token use, maybe I'm missing this 👀 I'm seeing more and more of our team moved to cowork, and lots of increased use, but a lot of ppl hitting token limits, which is great, but no clear way to track. I'm thinking maybe I can use Readout from @benjitaylor I'm thinking? I've just used it with Codex and Claude Code so not sure it works with Codex. Curious what other ppl are doing?

Looking to work with designers or agencies for an upcoming project for a new @terminal49 product. Share your portfolio and recommendations!

@LLMJunky @scottstts codex loves big bold rounded corners and cards

the problem is that this is the world's okayest website. and i mean that with absolutely 0 disrespect it is clearly ai slop but its not the worst thing i've ever seen. this honestly only proves the point more. that said, you can get some good UI's out of Codex Its just so much easier with Claude. much love keep grinding

Is gpt-5.4 actually bad at frontend design tho or do people just say it out of spite? I just asked it to make a one off reader tool for transcript inspection, and it one shot this It looks good and nothing like the vibe coded cliche

@kneath sounds personal

I’m trying hard to keep this new phase of social media focused on building and where I see the world going. But you know, long ago I worked with a16z as they tried to help me “fix” GitHub’s growth in 2012. There’s something you should know about Marc. He’s a fucking talentless tool. Ben is the one with talent. Here’s an outhouse (the bonehouse) I built this summer. I made it out of trees we milled on site with the help of Pat. Pat is a great guy. Fuck Marc Andreessen.

Got a @grok hat as a gift. Pretty impressed with the custom design and details. Does anyone know where to get something like this made?

@NickADobos #2 is fiction, might be real at some point but it's still a toy at the moment.

Current trajectory: 1. Codex app launches on iOS 2. People start replacing apps with prompts and skills 3. Entire iOS app ecosystem collapses 4. OpenAI builds a new phone device with codex as the only app 5. Apple becomes hardware only company

@abuiles Listening

You can your ask your claw, cursor, codex, claude code, or cowork agent to help you buy coffee, prompt this: Read lobsterbrew.com/skill.md and follow the instructions to join Lobster Brew 🦞☕️

@abuiles Or codex --yolo

codex: codex --sandbox danger-full-access --ask-for-approval never just codex 🚀

Complete the coding agent debug loop. Give your coding agent super powers so it can look a the dev server logs. Less copy and pasting, and more automated debugging. 1. I use github.com/DarthSim/overm… for dev process manager that uses tmux 2. Custom `just dev` command via github.com/casey/just 3. Agents.md instruction `just dev` runs through `overmind`. Agents should check for the current branch-scoped overmind socket before starting a new session and prefer `overmind -s status|echo|restart ` over probing tmux directly.

so good youtube.com/watch?v=iPZ5O8…

@0thernet @zocomputer @openclaw Awesome! Give your self a pat on the back, and make Zo awesome :) Until then I will continue to use openclaw

for the record, we see a lot of people saying @zocomputer is just hosted @openclaw... but really OpenClaw was inspired by Zo June 20, 2025 - we launch our beta Nov 19, 2025 - we launch 6 days later... first OpenClaw commit

@ashtilawat write your own words

Just finished reading through their paper. research.usecortex.ai/cortex.pdf The Good Stuff - Benchmark choice is legit - 90.79% vs 85.2% next best is a real gap, not noise - Sliding window enrichment (fixing the "orphaned pronoun" problem) is a solid contribution Where It's Weaker Than It Looks - They tested themselves on Gemini 3.0 Pro and competitors on GPT-4o? - The decay engine (the thing that stops graph rot) is openly "experimental" - Not sure what is driving gains... Sliding window? graph? - Append-only = no data loss sounds great, but it guarantees graph bloat over time - They deferred the maintenance problem, didn't solve it My Bottom Line - Sliding window enrichment → genuinely portable, bolt it onto RAG and get most of the gains - Temporal graph → only earns its complexity if your agent genuinely needs time-aware reasoning - Good hybrid RAG + contextual enrichment + timestamps gets you to ~85% without the graph maintenance headache - That last 5-6% probably isn't worth it for most use cases

@r_vonhagn @scaling01 @basedjensen Nothing really. Zero-sum thinking is a waste of time. Both BYD and Tesla can continue to win. Waymos are great. So is Tesla FSD. I use Waymo and have two Teslas with FSD subscriptions. They’re all great and have different use cases. Go build something :)

@dodeja @scaling01 @basedjensen Sorry, whats bringing down Elon? Companies like BYD and Waymo, or twitter conversations?

Elon is fumbling on multiple fronts

@zeeg yeah too much false group think

I for one am glad to see another loud voice of reason

Groupthink is getting worse in the AI era.

Folks are missing an important detail about MCP servers: distribution. Currently, they are the only standard way to auth to a service. To get your auth'd service on ChatGPT or Claude, you need to create an MCP with OAuth 2.1 support.

@scaling01 @basedjensen And you have invented...? What intellectual discourse are you having by bringing him or other entrepreneurs down?

@basedjensen he's losing everywhere Tesla autopilot is worse than Waymo Tesla cars are worse than Chinese EVs SpaceX is still doing fine with Falcon but Heavy is meh and starship is failing every test flight Boring company is obviously just a joke

@bri4nr33d @Austen I think there will be a huge demand for this but the pricing will collapse over time. So the TAM $ will shrink.

Okta CEO McKinnon Q3 FY26: Agent identity TAM "bigger than workforce ($50B) + customer identity ($30B) combined." Auth0 for AI Agents is GA — 100+ customers engaged representing $200M+ existing ARR, first deal closed (finserv, "significant ACV uplift"). Per-agent pricing model introduced.

Who is building the better auth layer for agents? Because what we have currently is miserable