typakon4 (@typakon4) header
typakon4 (@typakon4) profile picture

shitposter/peakposter & degen(erat) web3/ai-powered apps builder vibecoding & AI/ML engineering

Joined January 2019
Photo by typakon4 (@typakon4)Photo by typakon4 (@typakon4)Photo by typakon4 (@typakon4)Photo by typakon4 (@typakon4)Photo by typakon4 (@typakon4)Photo by typakon4 (@typakon4)Photo by typakon4 (@typakon4)Photo by typakon4 (@typakon4)Photo by typakon4 (@typakon4)
Epic built the Unreal Engine 5.9 Preview with a swarm of AI agents. And now your coding agent can work INSIDE the Unreal Editor. 🤯 A small team directed the agents to port AI tooling from UE6 and implement experimental browser support. The preview lets you run: • Claude Code • Codex • Cursor • OpenCode directly in the editor, connected to Unreal MCP. AI agents helped build the update that puts AI agents inside the editor. Game dev is about to get weird.
The Unreal Engine 5.9 MegaJam Preview is here! It’s the first Unreal Engine release built primarily with agentic workflows and opens up a few new Experimental features: check out Unreal MCP’s expanded toolsets, work with AI coding agents directly in Editor, explore AI-assisted content creation, and run your projects in a browser. This isn’t your typical Unreal Engine Preview, but it’s ready just in time for the Epic MegaJam. Invite your agents and start experimenting: epic.gm/ue-5-9-preview
7
14
230
Anthropic went from ~2,500 to 30,000+ code changes merged per day in 4 months. And the way they're using Claude Code now is kinda insane. Lydia Hallie from the Claude Code team says they're running dozens of sessions in parallel — for hours. The agents check their own work. Some orchestrate other agents. The shift: from steering one assistant to building the system that lets a whole fleet work.
3
16
478
Holy sh*t. Anthropic engineers built an AI research team on stage: → one agent researches company filings → specialists analyze financials and risk → another agent grades the result Then they show you how to build it. Free workshop: 4 videos + the code ↓
2
9
455
typakon4 retweeted
OpenAI ran a Codex agent for 1,000+ hours. In 6 weeks, it: • merged 213+ fixes • ran 5,800+ subagent sessions • sent 600+ Slack messages to coworkers Its job: debug failing tests and keep CI healthy. Est. cost: ~$32/hour, plus CI and review.
2
1
9
391
OpenAI ran a Codex agent for 1,000+ hours. In 6 weeks, it: • merged 213+ fixes • ran 5,800+ subagent sessions • sent 600+ Slack messages to coworkers Its job: debug failing tests and keep CI healthy. Est. cost: ~$32/hour, plus CI and review.
2
1
9
391
An Anthropic engineer gave Claude Fable 5.2 ONE prompt. 8 hours later, it had built a playable game. Felix Rieseberg says there was no human feedback during the run. The models, textures and sound? All created from scratch. But “one prompt” doesn't mean “cheap.” Here's what the run took: → ~445 million input tokens, heavily cached → ~2.7 million output tokens → ~$550 at API pricing This wasn't a quick chatbot reply. It was a model working on one task for hours. A prompt is starting to look less like a request for an answer and more like the beginning of a project. The question is whether the result is worth the compute. Would you spend $550 to build a game like this?
3
85
typakon4 retweeted
Claude Fable 5.2 just turned the history of science into a 5-minute film Derya Unutmaz says the 3D journey goes from early discoveries to a future with superintelligence - and was created entirely with code Here's one part:
1
4
190
Claude Fable 5.2 just turned the history of science into a 5-minute film Derya Unutmaz says the 3D journey goes from early discoveries to a future with superintelligence - and was created entirely with code Here's one part:
1
4
190
typakon4 retweeted
I built a Grok Bot that watches my Grok Bot usage and tells me HOW TO SAVE After usage meters jumped overnight for thousands of people with nothing running, everyone was furious, so I built my own watchdog It’s called Saver Bot, and it shows: > Where my usage goes, broken down by bot, routine, day and activity > An animated dashboard with a 30-day trend > Spike days, looping routines and the bot eating my week > Saving tips ranked by impact No made-up token counts or fake forecasts, just real usage data The first check already paid off. One news digest routine was 28% of my week. And when everyone's meter jumped on Oct 31, my daily usage proved the spike wasn't me Install it and it: 1/ Reads your usage and when it resets 2/ Builds your dashboard 3/ Flags anything unusual 4/ Suggests up to 3 fixes, biggest first 5/ Applies the one you pick, with an undo And yes, the template is completely Free: x.ai/bot/HwU-TfXaEV_zha7HpkG…
My Grok Bot limits ran out in two days, and I got tired of it So I found a way out of this situation and compiled tips in a single article
67
23
7
331
45,097
Your Harness. You Define It. Meet Creator Mode in DeepSeek Harness. Turn ideas into plugins that you can use, from a design workspace to a sandbox game or a 3D LLM visualizer. What will you create?
760
2,122
960
25,392
2,553,501
typakon4 retweeted
6
797
28
8,052
72,986
[BOOT SEQUENCE COMPLETE] // LIVE Cyberpunk: Edgerunners 2 is out now, only on Netflix! 🦾 ⚡
205
1,952
435
14,060
648,508
Take a break ☔
41
746
40
5,302
137,743
typakon4 retweeted
INTRODUCING FREEDZ What if the words we write freely are the things that imprison us? FREEDZ is an 8,888-piece deflationary PFP collection exploring freedom, control, and the machines shaping our world. A hidden world. A sacrificed word. A question. Are we really free? Here's the story:
525
3,982
94
4,749
93,554
typakon4 retweeted
your AI knows you. its competitor thinks you're a stranger. that's a switching cost no pricing page shows. switching models is one decision. rebuilding the context that makes an assistant useful is a different job. what you're working on. what you already know. which explanations waste your time. which assumptions it should stop making. if that understanding stays trapped inside one app, better personalization also makes leaving harder. this is the part of @conso_xyz's Consumer Passport direction that interests me. its Agentic Memory docs describe CPAC: connected-platform signals, including GitHub contributions, on-chain activity and AI usage, structured into portable context an agent can access with permission. the distinction matters: the docs describe structured reputation signals and context, not a dump of raw private conversations. this isn't the same as exporting a chatbot's existing memory. but moving a profile is only half the problem. a contribution graph doesn't explain what you're stuck on. a wallet history doesn't explain what you want next. a reputation score doesn't tell an agent when to stop explaining the basics. the hard part is turning those signals into something that actually changes the help you get. here's the comparison i'd want to see: same model. same task. with and without CPAC context. - fewer questions you've already answered elsewhere? - fewer wrong assumptions about your experience? - a more relevant first answer? and before trusting it: let me inspect what the profile says about me, correct a bad inference, and choose what an agent receives. those are the standards i'd judge it by. i haven't tested CPAC live; i'm discussing the direction in its docs. the reason to care goes beyond a collectible: a useful identity layer could let an agent earn your time through the quality of its work, instead of keeping you because starting over is exhausting. i want to choose the best agent for the job. not the one that already has my backstory. CONSO BOTS creator-campaign entry ⇩
4
1
10
670
your AI knows you. its competitor thinks you're a stranger. that's a switching cost no pricing page shows. switching models is one decision. rebuilding the context that makes an assistant useful is a different job. what you're working on. what you already know. which explanations waste your time. which assumptions it should stop making. if that understanding stays trapped inside one app, better personalization also makes leaving harder. this is the part of @conso_xyz's Consumer Passport direction that interests me. its Agentic Memory docs describe CPAC: connected-platform signals, including GitHub contributions, on-chain activity and AI usage, structured into portable context an agent can access with permission. the distinction matters: the docs describe structured reputation signals and context, not a dump of raw private conversations. this isn't the same as exporting a chatbot's existing memory. but moving a profile is only half the problem. a contribution graph doesn't explain what you're stuck on. a wallet history doesn't explain what you want next. a reputation score doesn't tell an agent when to stop explaining the basics. the hard part is turning those signals into something that actually changes the help you get. here's the comparison i'd want to see: same model. same task. with and without CPAC context. - fewer questions you've already answered elsewhere? - fewer wrong assumptions about your experience? - a more relevant first answer? and before trusting it: let me inspect what the profile says about me, correct a bad inference, and choose what an agent receives. those are the standards i'd judge it by. i haven't tested CPAC live; i'm discussing the direction in its docs. the reason to care goes beyond a collectible: a useful identity layer could let an agent earn your time through the quality of its work, instead of keeping you because starting over is exhausting. i want to choose the best agent for the job. not the one that already has my backstory. CONSO BOTS creator-campaign entry ⇩
4
1
10
670