I fully support Claude’s new initiative - why let competitors profit at their expense while losing users?

In theory, if things had continued as they were, it could have resulted in a loss of hundreds of millions in revenue.

As a founder and the creator of the company, I completely understand them. Their goal is to make money and retain users.

One person at X (link) is trying to convince me that Claude should have simplified integrations and then changed its business model, but I think he failed to consider that this would place a very heavy load on the infrastructure and, as a result, lead to financial losses.

Mozila found and dixed 271 security bugs in Firefox using AI

Mozilla shared how they used Claude Mythos Preview and other models to find an unprecedented number of latent security vulnerabilities in Firefox

What happened:

In April 2026, Mozilla shipped fixes for 423 security bugs total, 271 of which were found by Claude Mythos Preview:

- 180 high secured
- 80 moderate severity
- 11 low severity

How it works:

AI creates and runs test cases, reproduces bugs dynamically, and writes detailed reports. Mozilla built a full pipeline around it:

1/ target selection
2/ deduplication
3/ triage, patching

Swapping in a new model is trivial once the pipeline is in place

Notable finds:

- Sandbox escapes via IPC and race conditions
- A 20-year-old XSLT use-after-free
- Bypass of RLBox (in-process sandbox for third-party libraries)
- A rowspan=0 overflow of a 16-bit layout bitfield that fuzzers missed for years

Any project can start using an AI harness to find vulnerabilities today, even with simple prompts. The models are capable enough, and the techniques to filter noise are well understood

source: https://hacks.mozilla.org/2026/05/behind-the-scenes-hardening-firefox/

still don't understand people who post fake revenue photo/screenshots, like "my bot made $X" or "my app hit $X MRR"

you can literally see the dashboard was generated with one AI prompt like: 'generate a demo dashboard showing $3k in revenue'

instead of faking it, they could actually be building something real in a specific niche. the only people who fall for this are beginners - anyone with actual experience can tell immediately that this person has nothing behind the screenshot

Andrej Karpathy has joined Anthropic

yesterday, one of the most recognized AI researchers in the world announced he's moving to Anthropic

his career path:

> 2015 - 2017: co-founder and researcher at OpenAI
> 2017 - 2022: director of AI at Tesla (Autopilot, FSD)
> 2023: returned to OpenAI for one year
> 2024 - 2026: founded Eureka Labs (AI education)
> may 19, 2026: joined Anthropic

what he'll be doing:

working on the pre-training team under Nick Joseph. he'll be building a new team focused on using Claude to accelerate pre-training research. in other words, Karpathy will be teaching Claude to improve itself

why this matters:

this isn't just a hire - It's a signal!

Karpathy is one of the few people on the planet who bridges the gap between LLM theory and large-scale training in practice

and he's not the only one. over the past few months, CTOs from Instagram, Workday, Box, You .com and others have all joined Anthropic. not as leaders, as individual researchers. people are leaving CTO positions at billion-dollar companies just to do research at Anthropic

his words:

"I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time."

more to read:

1. OpenAI Cofounder Andrej Karpathy Joins Anthropic as Sam Altman’s Fortunes Turn
2. OpenAI co-founder Andrej Karpathy joins Anthropic

Kimi AI completely broke on a VERY simple prompt

I've been learning Russian for a while. since I have an architectural background, I decided to test Kimi with something basic:

a simple heating unit calculation with real pipe diameters, lengths, fittings, etc..

prompt I sent (in russian):

calculate the heat distribution for these sections:

section 1: 50mm diameter, 4.1m long, with turn tee section 2: 40mm, 1.45m, with tee and so on..

what I got back:

instead of any calculation, Kimi started endlessly repeating "I'll do this in all rubles, in all dollars, in all euros.." (translation) and listed literally every currency in the world in a crazy loop for dozens of lines

It was an endless cycle, maybe something wrong with the code

it kept going for 2 hours until I manually stopped it

the weird part:

it only burned 18% of my total tokens, so it wasn't even doing real computation, just looping text

details: https://x.com/0x_kaize/status/2059236004863631428

CREAO AI WAS HACKED

The team shared on their Discord that their X was hacked. The attackers then made a post about launching some token (don't fall for it).

Sharing this news since I'm their official partner:
https://x.com/CreaoAI/status/2059689476113514755

Chinese labs did it again - GLM 5.2 is crushing the competition

On June 13, Z.ai dropped GLM 5.2, and it's the first open-weights model that genuinely breathes down Claude's neck and beats GPT-5.5 on coding.

numbers that are hard to ignore:

1. SWE-bench Pro: 62.1 - decisively beating GPT-5.5 (58.6) and its own predecessor GLM-5.1 (58.4) BuildShip
2. it trails Claude Opus 4.8 by just 0.7 points on FrontierSWE (75.1 vs 74.4) and 0.8 on MCP Atlas YouTube
3. On Design Arena GLM-5.2 took first place, beating even Claude Fable 5 with an Elo of 1360 BuildShip
4. First open-weights model to cross 80% on Terminal-Bench BuildShip

the brutal part - the price:

$4.40 per 1M output tokens vs $25 on Opus 4.8 - 5.7x cheaper. and 6x cheaper than GPT-5.5. plus MIT license - you can download the weights and self-host.

basically frontier-level coding for pennies. Cline literally called it a "game changer" and said "open weights is back."

BUT there's a problem!

almost everyone who started using GLM 5.2 ran into the same wall - limits running out in an hour, because most people don't understand how the billing works (the Coding Plan counts prompts, not tokens).

I was burning through everything too fast myself, until I changed 10 things.

now I code all day and almost never hit the limit.

wrote it all up in detail - the caching trick that gives an 81% discount, the free models, self-hosting, the exact Claude Code config, and 6 more habits:

https://x.com/0x_kaize/status/2068775813785506091

Polymarket traders just bailed on GPT-5.6 - the "sure thing" bet cratered to 2%

a couple weeks ago a june release looked locked in. now the market has basically written it off.

the numbers:

1. the bet that GPT-5.6 ships in the june 22-28 window: 83% earlier this month → 2% now
2. the inverse - "not released by june 28" - now sits at 98%
3. $835k wagered on timing alone - one of polymarket's busiest tech markets right now

what's reportedly inside (nothing official from openai yet):

- 1.5M token context - ~43% jump over 5.5
- sharper UI + frontend code generation
- faster codex, stronger long-horizon agentic runs
- a rebuilt reward pipeline to kill the 5.5 reward-hacking bug

BUT

openai's been in IPO quiet period since its june 8 s-1 filing. so whenever 5.6 actually lands, it ships as a silent technical update - not a keynote.

the model looks basically done. the market just stopped pretending it knows when.

Trump allegedly died of rabies after being bitten by Vance

this isn't a tabloid headline - this is what DuckDuckGo's AI search was confidently telling users as fact.

what happened:

Reddit users deliberately spread a fake story across subreddits to test how easily AI search can be manipulated, the result exceeded expectations.

DuckDuckGo AI started confidently reporting that the president is dead. added on its own that Vance also died. and that RFK Jr. recommended treating Trump with more bites - because it could give him superpowers.

one coordinated Reddit post > AI picked it up > served it as verified information to millions.

next time an AI answers you with absolute confidence - remember this!

Recommend open source project - Claude Code From Scratch

This is an e-book on learning Claude Code, but it’s more than just a book: it comes with fully functional code, so you don’t have to wade through 500,000 lines of Claude Code to understand what’s going on.

~4300 lines (TypeScript and Python) recreating Claude Code's core architecture:

- Agent Loop
- 13 tools (parallel execution + early streaming start)
- 4-layer context compression
- Semantic memory recall
- Skills system
- Multi-agent
- MCP integration

Every step is mapped against the actual source code -here's how the team built it, and here's how we simplified it.

The book is split into 13 chapters.

Each one a hands-on tutorial:

You follow along, write a few thousand lines of code yourself, and walk away with a real sense of how the best coding agent on the market actually works under the hood.

I ran through it myself and already picked up things I didn't know.

Link: https://x.com/0x_kaize/status/2071955614561972566

Loop Engineering - the new discipline that blew up AI

Within a single week, three people from the industry said the same thing:

Peter Steinberger (creator of OpenClaw): "you shouldn't be prompting coding agents anymore - you should be designing loops that prompt your agents"

Boris Cherny (creator of Claude Code): "I don't prompt Claude anymore. I have loops running that prompt Claude. my job is to write loops"

Addy Osmani (Google Chrome) gave it a name: Loop Engineering

A week later Andrew Ng placed it all into the bigger picture of building products from zero:

The core idea is simple: you're no longer inside the loop (handing tasks to your agent every morning) - you step outside the loop and design it.

I wrote a full guide on building your own programmable agent from scratch:

- it wakes up on its own via cron
- finds work on its own (CI, issues, commits)
- criticizes its own code (a second reviewer agent that says "no")
- saves results to disk on its own
- schedules its own next run

5 lessons, all the code included, works on any project + the 5 loop diseases, the 4 debts that quietly pile up, and why without budget caps one bug can spin all night.

Link to article: https://x.com/0x_kaize/status/2073438517775003671

The complete GPT-5.6 guide

Sol, Terra, and Luna - three models, one generation, and the strangest launch in openai history (the one that went through a US government review first).

Inside:

1. What each tier actually is and who it's for
2. The benchmark openai wins big - and the one they'd rather you didn't look at
3. Ultra mode - multi-agent orchestration as a toggle, 4 sub-agents in parallel
4. The caching trick that cuts agent costs by 87%
5. Routing rules + ready-to-use api code

read it here: https://x.com/0x_kaize/status/2076290688124100986

One guy is pushing that this opencode config as the fix that "saves thousands of GPT-5.6 Sol tokens":

{
"$schema": "https://opencode.ai/config.json",
"permission": {
"task": "deny"
},
"tool_output": {
"max_bytes": 8192
}
}

That's partly true, but not entirely - you need to understand what you're giving up:

task: deny fully disables subagent spawning. It kills the recursion problem where subagents spawn subagents and tokens burn exponentially. But it also nukes every orchestrator workflow you might rely on. Fine if you don't use them, amputation if you do.

tool_output.max_bytes: 8192 caps tool output at 8 KB (default is 50 KB). Saves tokens on bash/grep/read returns. But on large outputs the model gets truncated context and often retries - burning the tokens right back.

the config works, BUT It's not a lifehack - it's two aggressive tradeoffs, fine if you know what you're giving up.

If you're worried that AI will replace you, read this post - especially if you work a 9-to-5 job.

A roadmap to make sure you never run out of money.

Based on my friend's experience, it took him only 3 months to learn AI, and he made the full transition to freelancing.

https://x.com/AlexFinn/status/2079377591522463813?s=20

Claude just gave me $100 in their tokens for free

Go to the website and check if you have this offer

You Can Pay $0 for Every Subscription

ChatGPT Plus $20/mo. Claude Pro $20/mo. Netflix $23/mo. Spotify $12/mo. Adobe $60/mo. Add 3 more and you're at $185/mo - that's $2,220/year for tools most people use 30% of

I wrote a full guide with 25 verified methods for 2026:

- the Cancellation Retention Trick (30-50% off in one click)
- student Discounts (free or half-price with .edu)
- referral Loopholes (invite yourself, get free months)
- annual Billing (15-20% off instantly)
- regional Pricing (Turkey/Argentina/India tricks)
- free Alternatives (replace $20/mo with $0)
- the Free AI Stack (25 free frontier models)

example of what's inside:

before applying tricks: $185/mo ($2,220/year)
after applying tricks: $69/mo ($828/year)

savings: $1,392/year with 30 minutes of setup.

full article on X: https://x.com/0x_kaize/status/2086490566171230403

You can get Cursor Ultra Plan for FREE

https://x.com/0x_kaize/status/2087620115378397501

You can get Cursor Ultra, SuperGrok Manus 1.6, Flux-3 and many other for FREE

I've put together a large collection of ways to get these tools for Free:

https://x.com/0x_kaize/status/2087993368043307294