Sumplex

AI enjoyer and builder

[Sumplex @0xSumplex]

Jev founder Diogo Almeida, formerly at OpenAI:

“Claude Code and Codex are still assistance tools.

Jev is what comes next beyond the human-in-the-loop era.”

LLMs → Assistance → Jev → Autonomous Systems

In this 140-minute technical talk, Diogo breaks down why RLHF is losing relevance and how Jev enables reliable autonomous decisions:

00:02:50 – What Jev really is
00:05:54 – Why RLHF produces confident mistakes
00:24:59 – RLCD vs RLHF and RLVR
00:39:55 – Building reliability, robustness, and determinism

Choice → Score → Noul → Action

Easily worth more than most $1,500 AI engineering courses.

Bookmark it and watch today.

Then read the complete guide to becoming a Jev engineer below ↓

[rari @0xwhrrari]

Most AI agents waste tokens on decisions that never needed text

Jev turns routing, scoring, and verification into a fast decision layer

I broke down the architecture most agent builders are still missing ↓ https://t.co/02POtifVJx

[Sumplex @0xSumplex]

Creator of GrokBot (SpaceXAI):

“99% of people use only 1% of GrokBot’s real power. They run one or two agents without a ‘loop’ or ‘graph’

I’m running 15+ GrokBot agents as one coordinated team. It has a Chief of Staff bot, 2 manager agents, and 13+ workers — this is the new engineering stack”

In this 1-hour podcast, a SpaceXAI engineer explains how to unlock 100% of GrokBot’s real potential

This is worth more than another $500 agentic engineering course

Watch the workshop today, then read the article below to learn how to build your own team of GrokBot agents

[Ichigo @iiiichigo_chan]

https://t.co/Btb2P1NieM

[Sumplex @0xSumplex]

Jev founder Diogo Almeida, formerly at OpenAI:

“The next era won’t belong to Claude Code or Codex

They’re still part of the assistance era, where a human remains in the loop

JEV is the next step for LLMs

220x faster, 410x cheaper, zero hallucinations, and no human in the loop”

LLMs → Assistance → JEV → Autonomous Systems

In this 37-minute technical talk, Diogo breaks down why RLHF is becoming obsolete and how the next generation of LLMs will be built

Train → Verify → Execute → Improve

This talk is worth more than a $200,000 Stanford Machine Learning degree

Bookmark it and watch it tonight

Then read the article below

[rari @0xwhrrari]

Most AI agents waste tokens on decisions that never needed text

Jev turns routing, scoring, and verification into a fast decision layer

I broke down the architecture most agent builders are still missing ↓ https://t.co/02POtifVJx

[beamnxw ./ @beamnxw]

I BUILT AN AUTONOMOUS SYSTEM WHERE GPT-6 ASTRA AND JEV BEAT MINECRAFT. LIVE RIGHT NOW AT https://t.co/rS42NQYfjO

Minelog. GPT-6 Astra and Jev in one Minecraft body. A live HUD. A public harness

I spent the last week on this. @ValsAI started pushing Minecraft as the test for agents and that idea sat in my head. I took those two models, trained the system for 144 hours, ran terabytes of traces, and wired the whole loop into a repo you can read

How the build is laid out:

1\ Astra is the slow mind. It sits in the background and writes a JSON plan: the stage, the goal, the waypoint, the items the body needs, the skill names the harness should lean on. The tick keeps moving while Astra thinks

2\ Jev is the fast mind. TypeSafe System One. Every tick it gets a compact snapshot of the world. HP, hunger, inventory, light, hostiles, nearby blocks, the current plan. Then it answers three typed questions. Which action. How safe. How urgent. One move from a closed catalog. Punch. Craft. Smelt. Mine. Walk. Eat. Sleep. Attack. Place

> The harness runs the tick
> Observe the world into JSON
> Reflexes in code for lava, drowning, creepers, death
> Astra in the background
> Jev on the hot path
> Policy gates the action
> The body executes through the player protocol
> A line lands in runtime .log

That log is what the site reads

Offical account: @minelogxyz
Offical website: https://t.co/xwv6xWrhQk

Gameplay on the left. The two minds on the right. Day, advancements, deaths under it

The campaign is eight stages. Wood. Iron. Diamonds. Nether. Fortress. Eyes. Stronghold. Dragon. Deaths stay in the tape. Recovery is a skill. A run ends when the dragon is down and the body walks the exit portal

https://t.co/WyDwnkZqm2
loop, planner, controller, skills, protocol, log, verify

⭣ ⭣ ⭣

[Sumplex @0xSumplex]

Anthropic engineer:

“99% of people still use Claude Code like Google. The other 1% are building harnesses that coordinate swarms of self-improving Claude agents

I run 100+ agents through a single harness, with Chief agents and PM agents orchestrating the entire team”

In this 30-minute workshop, an Anthropic engineer shows how to get maximum value from Claude Code while keeping costs to a minimum

Worth more than yet another $500 vibe-coding course

Watch it today, then read the article below to learn how to build your own self-improving agent harness

[Ichigo @iiiichigo_chan]

https://t.co/Esm9l92Hkb

[Sumplex @0xSumplex]

this is pure f*cking leverage

10 AI engineering resources worth keeping one click away

agents. harnesses. context. tools. evals. tracing. production patterns.

01 OpenAI Agents SDK
▸ https://t.co/8iIPtYCKrZ

02 Anthropic Context Engineering
▸ https://t.co/HxHC1QeFUJ

03 Model Context Protocol
▸ https://t.co/CFndAZtv88

04 Google Agent Development Kit
▸ https://t.co/1awdJUotqQ

05 LangGraph
▸ https://t.co/9759jJorsK

06 Testing Agents with PydanticAI
▸ https://t.co/kLtry0A3et

07 OpenAI Evals Best Practices
▸ https://t.co/dg0sMpc9EH

08 Arize Phoenix
▸ https://t.co/4bMOfUw04k

09 SWE-bench
▸ https://t.co/eweKVM8K5E

10 Hugging Face Agents Course
▸ https://t.co/vmg7FPadFZ

the loop:

learn the primitives → build the harness → connect tools → add state → trace every run → evaluate failures → ship → repeat

save this before your next “agent” becomes a 2,000-line prompt with zero tests ↓

[Sumplex @0xSumplex]

this is pure f*cking leverage

10 AI engineering resources worth keeping one click away

agents. harnesses. context. tools. evals. tracing. production patterns.

01 OpenAI Agents SDK
▸ https://t.co/8iIPtYCKrZ

02 Anthropic Context Engineering
▸ https://t.co/HxHC1QeFUJ

03 Model Context Protocol
▸ https://t.co/CFndAZtv88

04 Google Agent Development Kit
▸ https://t.co/1awdJUotqQ

05 LangGraph
▸ https://t.co/9759jJorsK

06 Testing Agents with PydanticAI
▸ https://t.co/kLtry0A3et

07 OpenAI Evals Best Practices
▸ https://t.co/dg0sMpc9EH

08 Arize Phoenix
▸ https://t.co/4bMOfUw04k

09 SWE-bench
▸ https://t.co/eweKVM8K5E

10 Hugging Face Agents Course
▸ https://t.co/vmg7FPadFZ

the loop:

learn the primitives → build the harness → connect tools → add state → trace every run → evaluate failures → ship → repeat

save this before your next “agent” becomes a 2,000-line prompt with zero tests ↓

[Ichigo @iiiichigo_chan]

https://t.co/Esm9l92Hkb

[Sumplex @0xSumplex]

Google engineer:

“Build a harness that runs self-improving agentic workflows for you around the clock with loops and graphs - this is the new job!

At Google, 90% of engineers already run dozens of agentic workflows inside their harnesses. This is what the next era of engineering looks like”

In this 1-hour conversation, a Google engineer explains what the future of AI engineering will actually look like

Worth more than 10 paid agentic courses

Watch it today, then read the article below to learn how to build your own self-improving agent harness

[Ichigo @iiiichigo_chan]

https://t.co/Esm9l92Hkb

[Sumplex @0xSumplex]

This paper is f*cking insane.

Prompt Engineering just got replaced by Harness Engineering.

A new 7-page paper held GPT-5 fixed and changed only the system around it:

Codex agent → 71.75%

Recursive Agent Harness → 81.36%

No larger model.

No fine-tuning.

A parent agent writes code that spawns full sub-agent harnesses in parallel.

Each gets its own context, filesystem, tools, and planning loop.

The authors tested it on 199 tasks with contexts reaching 4M tokens.

With Claude Sonnet 4.5, the same harness reaches 89.77%.

The model is no longer the whole system.

The harness around it is.

Bookmark this, then read the full Harness Engineering guide below ↓

[Sumplex @0xSumplex]

This paper is f*cking insane.

Prompt Engineering just got replaced by Harness Engineering.

A new 7-page paper held GPT-5 fixed and changed only the system around it:

Codex agent → 71.75%

Recursive Agent Harness → 81.36%

No larger model.

No fine-tuning.

A parent agent writes code that spawns full sub-agent harnesses in parallel.

Each gets its own context, filesystem, tools, and planning loop.

The authors tested it on 199 tasks with contexts reaching 4M tokens.

With Claude Sonnet 4.5, the same harness reaches 89.77%.

The model is no longer the whole system.

The harness around it is.

Bookmark this, then read the full Harness Engineering guide below ↓

[Ichigo @iiiichigo_chan]

https://t.co/Esm9l92Hkb

[Sumplex @0xSumplex]

Andrej Karpathy:

“Prompting is fading away.

The harness around the model is becoming the real product.”

In a 1-hour Stanford lecture, he explains why the model alone is never the complete system.

The missing layer most people overlook:

Model → Harness → Feedback → Working Software

The model provides intelligence.

The harness turns it into reliable output.

Watch it first.

Then read the full guide on Harness Engineering below ↓

[Sumplex @0xSumplex]

Andrej Karpathy:

“Prompting is fading away.

The harness around the model is becoming the real product.”

In a 1-hour Stanford lecture, he explains why the model alone is never the complete system.

The missing layer most people overlook:

Model → Harness → Feedback → Working Software

The model provides intelligence.

The harness turns it into reliable output.

Watch it first.

Then read the full guide on Harness Engineering below ↓

[Ichigo @iiiichigo_chan]

https://t.co/Esm9l92Hkb

[Sumplex @0xSumplex]

Andrej Karpathy spent 8 years at OpenAI and Tesla.

He packed everything he knows about LLMs into one free 2-hour lecture.

Models → Context → Tools → Agents → Harnesses

People pay $15k for bootcamps that teach less than this.

By the end, you’ll understand why the model is only one part of the system—and why the harness around it matters.

You probably don’t have 2 hours right now.

Don’t let this disappear from your feed.

Watch it, then read the Harness Engineering guide below and build the system around your first agent.

[Sumplex @0xSumplex]

Andrej Karpathy spent 8 years at OpenAI and Tesla.

He packed everything he knows about LLMs into one free 2-hour lecture.

Models → Context → Tools → Agents → Harnesses

People pay $15k for bootcamps that teach less than this.

By the end, you’ll understand why the model is only one part of the system—and why the harness around it matters.

You probably don’t have 2 hours right now.

Don’t let this disappear from your feed.

Watch it, then read the Harness Engineering guide below and build the system around your first agent.

[Ichigo @iiiichigo_chan]

https://t.co/Esm9l92Hkb

[Sumplex @0xSumplex]

Anthropic engineer:

“90% of our engineers were already running self-improving loops.

Now everyone is building the harness around them.”

“No more prompt-only workflows.”

In just 10 minutes, she builds her complete Claude Code setup live from an empty terminal.

Model → Harness → Reliable Software

The prompt is only one component.

Harness Engineering is the system around it.

This 10-minute video is worth more than most $1,000 agent engineering courses.

Watch it today.

Then save the full Harness Engineering guide below before everyone catches up ↓

[Sumplex @0xSumplex]

Anthropic engineer:

“90% of our engineers were already running self-improving loops.

Now everyone is building the harness around them.”

“No more prompt-only workflows.”

In just 10 minutes, she builds her complete Claude Code setup live from an empty terminal.

Model → Harness → Reliable Software

The prompt is only one component.

Harness Engineering is the system around it.

This 10-minute video is worth more than most $1,000 agent engineering courses.

Watch it today.

Then save the full Harness Engineering guide below before everyone catches up ↓

[Ichigo @iiiichigo_chan]

https://t.co/Esm9l92Hkb

[Sumplex @0xSumplex]

Anthropic engineer:

“At Anthropic, we don’t rely on prompts alone anymore

We engineer the harness around the model”

In 30 minutes, she shows how the Claude team builds the context, tools, tests, and feedback loops that keep agents working reliably

If this cost $400, people would call it the best agent engineering course of the year

It’s free

Watch it, then use the step-by-step guide below to build your first agent harness

[Sumplex @0xSumplex]

Anthropic engineer:

“At Anthropic, we don’t rely on prompts alone anymore

We engineer the harness around the model”

In 30 minutes, she shows how the Claude team builds the context, tools, tests, and feedback loops that keep agents working reliably

If this cost $400, people would call it the best agent engineering course of the year

It’s free

Watch it, then use the step-by-step guide below to build your first agent harness

[Ichigo @iiiichigo_chan]

https://t.co/Esm9l92Hkb

[Sumplex @0xSumplex]

OpenAI CEO Sam Altman:

“You no longer need to write every prompt yourself anymore.”

In this 40-minute Stanford lecture, he explains how to use ChatGPT at a level most people never reach.

LLMs → Prompts → Agents → Harnesses

The harness supplies the context, tools, tests, and feedback that make it reliable.

He originally delivered this lecture to Stanford students.

A friend sent me the recording.

I watched it last night and realized I was using less than 15% of what ChatGPT can actually do.

Watch it, then read the full Harness Engineering guide below.

[Ichigo @iiiichigo_chan]

https://t.co/Esm9l92Hkb

[Sumplex @0xSumplex]

@iiiichigo_chan one of the best video by Karpathy