Moosa Jaffery | Building AI Agents.

Building Agentic AI Products @voir_labs | Web3, DeFi Products & Engineering @xorddotcom | prev @aisharktank, @a51_Fi, @spockanalytics,@ToDaMoonAI,@mesh_protocol

[Moosa Jaffery | Building AI Agents. @_MoosaRaza]

#Web3 ecosystem is facing a significant #challenge in driving #growth, causing many high-quality ideas to struggle to survive. ✈️
Thread 🧵
(1/9)

[Voir Labs @voir_labs]

The #1 AI model in the world was pulled offline 72 hours after launch.

Not hacked. Not recalled. Banned — via U.S. export controls.

The trigger: Amazon researchers typed "fix this code" and got vulnerability PoC outputs. Andy Jassy escalated to Commerce. BIS issued a directive. Fable 5 was gone.

126+ security leaders signed an open letter saying it wasn't a jailbreak. The only outside expert to review the Amazon report called it "a complete overreaction."

Meanwhile: OpenAI's Daybreak does the exact same flaw-finding. By design. At scale. With CrowdStrike, Cisco, and Palo Alto as partners.

It's not banned.

The difference isn't capability. Anthropic said directly that GPT-5.5 produces identical outputs. The difference is whose posture the current administration trusts.

Here's the thing nobody is saying plainly:

Safety positioning was never just ethics. It was a go-to-market strategy. It attracted enterprise buyers, alignment talent, and regulatory credibility.

It also made Anthropic the obvious target when the government wanted to set a precedent.

The labs that survive AI governance won't necessarily be the ones that built the safest systems. They'll be the ones who figured out how to make their safety story legible to people who can invoke BIS controls at a moment's notice.

That's a different skill than alignment research.

Anthropic has the latter. The Fable 5 ban is evidence they've underinvested in the former.

[Moosa Jaffery | Building AI Agents. @_MoosaRaza]

Modern alms 🦴

[ClaudeDevs @ClaudeDevs]

We've reset 5-hour and weekly rate limits for all users. Enjoy Fable 5!

[Moosa Jaffery | Building AI Agents. @_MoosaRaza]

@KashanSid3 @nikitabier +1

[Moosa Jaffery | Building AI Agents. @_MoosaRaza]

DeFi Reborn ✌🏻

[Kashan Siddiqui @KashanSid3]

Pressure doesn’t test the system, it tests the will of those who stand behind it.

The hack did put DeFi at risk, no denying that. But we will remember this day not for that, but because DeFi was born again.

Love this community. DefiUnited
Track live: https://t.co/UeSpFql7BU

[Voir Labs @voir_labs]

OpenAI just shipped GPT-5.5, the first fully retrained base model since GPT-4.5.

The benchmark headlines are real. Leads on 14 public benchmarks. 82.7% on Terminal-Bench 2.0. 84.9% on GDPval.

But here's the part most coverage is glossing over:
Claude Mythos beats it on SWE-bench Verified (93.9% vs 88.7%) and SWE-bench Pro (77.8% vs 58.6%).

The difference? Mythos isn't available. Invitation-only. No public API.

For builders who need to ship today, GPT-5.5 is the top model you can actually use.

A few things worth knowing before you update your routing:

1. The "half the cost" claim is real, but specific. Raw per-token pricing doubled vs GPT-5.4 ($5/$30 vs $2.50/$15). The efficiency story only holds for agentic Codex-style workloads, where it uses ~40% fewer output tokens. General inference? You're paying double.

2. GPT-5.5 launched the same day Anthropic published its quality degradation post-mortem. Three bugs between March 4 and April 16, one undetected for 34 days. OpenAI didn't engineer the timing, but the credibility window is real.

3. Codex used GPT-5.5 to rewrite OpenAI's own GPU load-balancing heuristics before launch. 20%+ token generation speed improvement. That's not a benchmark. That's production proof.

The availability gap is what actually moves product decisions.

[Voir Labs @voir_labs]

Anthropic quietly removed Claude Code from the Pro plan. Then reversed it 24 hours later. But the reversal isn't the story.

Agentic coding sessions run hours long. They burn through context windows, chain tool calls, and loop through terminal executions. A serious developer on Claude Code can consume in one session what a casual user does in a week.

That doesn't fit inside a $20/month flat plan. And Anthropic just made that problem visible — accidentally.

This isn't a Claude story. It's a category story. Every lab offering agentic coding inside a consumer subscription is carrying the same structural exposure.

Read the full article below!

@AnthropicAI @claudeai

#AI #Claude #Anthropic

[Moosa Jaffery | Building AI Agents. @_MoosaRaza]

@KashanSid3 💯

[Voir Labs @voir_labs]

https://t.co/2LZ3yP17A9

[Moosa Jaffery | Building AI Agents. @_MoosaRaza]

Another Hack on @solana : The Governance Failure🔓

@DriftProtocol hit for $285M is a grim reminder for builders.Listing fake collateral and disabling guards in one TX isn't just a hack, it’s an architectural failure

A 2/5 multisig with 0s timelock isn't DeFi; it's a liability.

[Moosa Jaffery | Building AI Agents. @_MoosaRaza]

one hundred percent 🙂

[Andrew Gazdecki @agazdecki]

POV: you realized distribution is more important than building the best product.

[Uttam @uttam_singhk]

never ask a woman her age, a man his salary, or a tech bro if his mac mini running OpenClaw has made any money yet.

[0xMarioNawfal @RoundtableSpace]

Anthropic dropped a 33 pages cheat sheet for building Claude skills

https://t.co/jEuH95NGn3

[Zac Glover @zacglover]

banger ad from @AnthropicAI

rather than fuel ppl’s fear, reframe the threat of ai in their favor.

[Moosa Jaffery | Building AI Agents. @_MoosaRaza]

average day in tech 😅

[mert @mert]

ceo: let's build product X

devs: hmm ok. that will require 3 months of work and we need to hire a few more engineers. also, I need to wrap up my current work.

ceo on socials: announcing product X! going live tomorrow.

devs:

[Alex Immerman @aleximm]

https://t.co/EZ4nTOUNo6

[Moosa Jaffery | Building AI Agents. @_MoosaRaza]

💀

[Uttam @uttam_singhk]

It sucks that claude is down but it *really* sucks for the brave soul who has to fix claude without using claude

[Moosa Jaffery | Building AI Agents. @_MoosaRaza]

👀

[Balaji @balajis]

We need AI market cap.

Meaning: Coinmarketcap, but for AI usage.

What’s the best coding model today?
The best image model?
Best video model?

It just constantly changes. So: rank on the basis of Google trends, usage by your friends, and other variables.

[CZ 🔶 BNB @cz_binance]

You need an AI to keep up with AI.

claw: what can I do you for today?

me: install every other bot, test it and keep me up-to-date on all new AI release going forward. Oh, and go wild respond to all my emails however you like.

[Moosa Jaffery | Building AI Agents. @_MoosaRaza]

@KashanSid3 @zachxbt 👀

[Voir Labs @voir_labs]

@AnthropicAI's latest on distillation detection highlights a growing divide in the AI industry. Is it about protecting IP or gatekeeping intelligence? 🛡️

The case for:
Frontier AI R&D costs billions. Distillation without consent is effectively IP theft that disincentivizes the next generation of models.

Anthropic alleges that labs like @deepseek_ai, @Kimi_Moonshot, and @MiniMax_AI used millions of queries to "steal" Claude's capabilities.

Anthropic’s detection research is a win. Builders deserve to protect their breakthroughs from predatory scraping. 🏗️

The case against:
But one can argue this is moat-building disguised as security research. "Preventing distillation" is just a polite way to gatekeep intelligence and force developers into permanent API dependency. Progress happens when knowledge is shared, not when it's locked behind a proprietary layer. 🔓

Is this a necessary security evolution or a blow to permissionless innovation?
The answer likely depends on whether you are building the infrastructure or the application layer.

#AI #Claude #DeepSeek