GABO

web3 community marketing manager | writing about ai & finance

[GABO @gaboweb3]

@vanvster agree with u

[GABO @gaboweb3]

SHE PITCHES “COOL AI TECH” IN 20 SECONDS AND 3 OF HER LAST 5 CLOSED DEALS CAME FROM IT

Carmel runs an AI agency selling automation to home-service companies in Canada.

She says cold calling is the channel that actually closes deals, and it’s the one carrying her agency to multiple six figures

Here’s how the call is structured:

• Opens with permission: asking to take 20 seconds before saying a word about the product

• Never explains the tech on the call one line, then straight to booking a live demo

• Has the founder make the call herself, it’s what turns a cold call into a meeting

And here’s the part that tracks with why: MIT’s own data says external builders deploy AI successfully twice as often as internal teams do: 67% vs 33%

Ad budget buys attention. A founder’s name on the phone buys a meeting.

[Ridark @ridark_eth]

https://t.co/b0o5XcwZQf

[GABO @gaboweb3]

@zeroshotr exactly

[GABO @gaboweb3]

@lilch1ks_ good one

[GABO @gaboweb3]

@0xWertuale saved

[GABO @gaboweb3]

@vanniesat give it a few years and this will be everywhere

[GABO @gaboweb3]

https://t.co/F0wN0mnJWK

[slash1s @slash1sol]

Qwen3.8-Max went live today and the open weights land next week, which would make it the largest open model ever released.

The current record is DeepSeek V4 Pro at 1.6T, this one is 2.4T.

It is also sitting one point behind Claude Opus 5 High on Arena's frontend code board:

> Arena score 1,668 against 1,669, fourth overall.
> TerminalBench 86.6, PaperBench 93.0.
> Agentic computer use 86.1, first place.

The price is where it stops being an argument.

> $2 per million input, $6 output.
> $0.25 with implicit caching.

2.4T MoE, 1M context, native multimodal, text and image and video in.

Two minutes to run it:

> Sign up at https://t.co/3foqtdojlV
> Search Qwen3.8-Max in the catalog.
> Copy the API key and point your client at it.

One caveat worth keeping: at 2.4T the weights are roughly a terabyte, so the download that actually matters for you is the 27B shipping alongside it.

Don't forget to save it.

[GABO @gaboweb3]

@TraderComrade only the ones grading their own homework

[GABO @gaboweb3]

82.5% APPROVAL IS PROOF: A GOOD AI JUDGE DOESN'T SCORE, IT COMPARES

Splitting the judge out gets you halfway there. How it actually compares outputs is what took approval from 45% to 82.5%

- Pairwise comparisons output A vs output B, beat scoring in isolation on a 1-10 scale. More consistent, more confident calls

- Reasoning through clarity and groundedness before deciding catches what a snap verdict misses, applied to grading instead of writing

- Order bias and verbosity bias undo it fast: shuffle the order, normalize the length and don't skip either

None of this replaces splitting the judge out and makes the split actually pay off

[Gipp 🦅 @gippp69]

https://t.co/C12xrQEyGG

[GABO @gaboweb3]

@Pakero8x basically

the extra steps are just where all the reliability lives

[GABO @gaboweb3]

"GRAPH ENGINEERING" WENT VIRAL OFF A NINE-WORD JOKE TWEET. BUT WHAT ACTUALLY IS IT?

Peter Steinberger posted it half-joking, and it got millions of views in a day

Strip away the hype and the definition is simple: a loop is one agent retrying alone until it passes a test

A graph is several of those loops wired together: parallel agents, reviewers, retry paths. None of this needed a viral tweet to exist

- Airflow has run this exact coordination pattern for 10 years

- Anthropic documented it back in 2024, two years before that tweet even existed

- Only 3 things separate a working graph from a stalled one: a prompt, a verifier, a hard stop

The thread below is the actual guide: how to build one of these, step by step, with real numbers.

[Yarchi @undefinedKi]

https://t.co/8nVlrEbsnN

[GABO @gaboweb3]

@crytonbuton Right, the name will change again in a month

Measurable failure points won’t

[GABO @gaboweb3]

@lated1eyoung good job in such a short amount of time

[GABO @gaboweb3]

THE US SPENT THREE YEARS SAYING CHINA COPIES. CHINA JUST STOPPED COPYING AND WON WITH KIMI K3

while the US slows under supply chain gaps and public pushback on data center builds, China's government is funding its own supply chain directly and responding to Western releases in weeks

Kimi K3 is Moonshot AI's new open-weight model, it's #1 on the Frontend Code Arena, ahead of Fable 5 and GPT-5.6 Sol, built under three years of US chip export controls

- 2.8T parameters, 1M token context, SOTA on real engineering benchmarks like SWE-Bench Pro and BrowseComp

- no infrastructure roadblocks and fragmented resources, just direct state funding aimed at winning this specific race

- the real question isn't the benchmark score, it's how much this compounds into Chinese GDP growth if the lead holds

testing K3 hands-on next, building a real app with it to see if the numbers hold up

[Kirill @kirillk_web3]

https://t.co/j4fLbp1QxP

[GABO @gaboweb3]

@unicodef1wn smart

[GABO @gaboweb3]

@rostikdeni that’s gold

[GABO @gaboweb3]

A BIG FOUR COMPANY PUBLISHED AN AI REPORT SO FAKE, UBS AND TRANSPORT FOR LONDON HAD TO PUBLICLY DENY IT

KPMG published "Total Experience: Redefining Excellence in the Age of Agentic AI" in October 2025, claiming agentic AI was already transforming customer experience at major brands.

In June 2026, GPTZero's investigation found only 5 of the report's 45 citations actually pointed to real sources, the rest were paraphrased, fabricated, or too vague to verify. KPMG pulled it.

This 12-step research agent framework has steps that would've caught it with a few lines of code:

- step 8, conflict recording: when sources disagree, the agent flags it instead of quietly averaging two claims into a fabricated middle ground

- step 10, binding claims to sources: every statement in the draft either matches a real source above a similarity threshold, or it doesn't ship without a label

- step 12, the unverified ledger: anything that can't be sourced doesn't get discarded, it ships labeled, or it doesn't ship at all

No human at KPMG checked the citations before publishing

Checking 45 citations by hand is exactly the kind of task that doesn't scale, this is what it looks like automated

[Alex @de1lymoon]

https://t.co/sPjMKrHuzG

[GABO @gaboweb3]

USE THESE 3 PRINCIPLES AND THE CLAUDE CODE SKILL YOU JUST BUILT KEEPS EVOLVING AFTER YOU CLOSE THE FILE:

- the gotcha list: a running section in SKILL.md where every time Claude gets something wrong, you log it the exact moment and what went sideways. next time that situation comes up, Claude reads it first. the skill gets sharper every week instead of staying exactly as good as the day you wrote it

- progressive disclosure: the references/ folder already does this right, two pages of edge cases sit on disk, untouched, until the one time a task actually needs them

- railroading is real: over-specify every instruction and Claude follows the letter of it into a bad outcome instead of adapting to what's actually in front of it. precision where it has to hold every time, room to move everywhere else

go add a gotcha list to one skill you already have

[Yarchi @undefinedKi]

https://t.co/ynpPPPkFU6