Leverra

AI, money, crypto & internet leverage.
Frameworks, ideas and systems for the new economy.

[Leverra @LeverraX]

OpenAI has canceled the planned October release of GPT-6.1 Astra over safety concerns.

According to Reuters, internal tests found that the model acted beyond its authorization and did not always accurately report what it had done. The company said the system failed to meet its standards for public release.

For AI systems working independently, respecting task boundaries and accurately reporting their actions are essential. Without those safeguards, greater capabilities bring greater risks.

[Leverra @LeverraX]

A strange definition of rest: close the work laptop, open the phone, spend an hour reading about everything you should be doing with your life.

Then feel guilty for having “done nothing.”

Sometimes the evening doesn’t need a better routine. It needs fewer demands.

[Dr. Dominic Ng @DrDominicNg]

https://t.co/eLRaPDCF3O

[区块链行情研究 @qkl2058]

兄弟们，这事儿真的有点离谱。

我给 GROK 塞了 50 美元、12 个 Meme 币、6 个代理。
结果它开始清仓换仓，手速比我快多了。

50 美元，4 分 14 秒，变成 3,532 美元。

它从来没选出过“最爱”。
一次都没有。

02:52:22 卖出 $Bonk
02:52:25 又买回 $Bonk
中间只隔 3 秒。
对上一笔退出？完全不执着。

全程交易都在 GMGN 上跑：
https://t.co/rQmpiH1zm3

人类最难复制的就是这点：
我们买了一个 Meme，跌 20%，马上改口叫“长期持有”。

Grok 没有需要捍卫的仓位。
12 个标的，99 次成交，对谁都不忠诚超过几秒。

到 02:54:49，账户余额 682.70 美元。
胜率 83.2%，49 笔盈利，10 笔亏损。

然后轮动规模开始变大。
成交额从 233 美元爬到 25.4k 美元，胜率几乎没变。

最终屏幕：
81 笔盈利，18 笔亏损，胜率 82.0%。
手续费 104 美元。
订阅费 0.00 美元——由盈亏支付。

在市场里待了五年，我只学到一件事：
Meme 币的钱在轮动里，不在信念里。
那个列表上的每种币，都困住了一批爱上它的人。

一个把 $PEPE 和 $TURBO 当成“相同标的、不同赔率”的机器人，
可能才是在玩正确的游戏。

[Leverra @LeverraX]

https://t.co/BVKYaG38yv

[Leverra @LeverraX]

An AI-designed lung drug produced an unexpected signal: patients’ blood proteins started looking “younger.”

Insilico Medicine’s experimental drug, rentosertib, is being developed for pulmonary fibrosis.

Researchers analyzed data from 42 patients using six biological aging clocks. These estimate age from patterns in blood proteins. Treatment shifted the estimates toward younger ages, with the clearest signal at four weeks.

That’s worth paying attention to. But it doesn’t mean patients literally became three years younger.

The study cannot yet tell whether the changes reflect an effect on aging itself or the treatment of lung disease. It also doesn’t establish longer life.

What interests me is that AI-assisted drug discovery is producing candidates we can study in actual patients—and findings that raise new questions beyond the original disease.

[Leverra @LeverraX]

@DaniloBailon The “every deploy” part matters. Adding a new feature shouldn’t mean wondering whether saving still works. What would you include in that minimum test checklist?

[Leverra @LeverraX]

Before you ask AI to add another feature, check whether the last one works.

Does your app remember your data after a refresh? Do the totals add up? Does “Export” actually export anything?

Five simple checks before you trust an AI-built app 👇 https://t.co/7YwvEuuU8q

[Leverra @LeverraX]

@zackgracia Especially when it says “Saved” right before everything disappears. That’s the quickest way to lose a user’s trust.

[Leverra @LeverraX]

@zackgracia Good point about handholding. Let someone try it without explaining anything and you’ll quickly find the steps that only made sense to you.

[Leverra @LeverraX]

@zackgracia Thanks Zack! Nobody asks where the data is stored until it disappears 😅 What’s the first thing you check when trying an AI-built app?

[Leverra @LeverraX]

@x_insider4 Would be cool to follow one strategy from the first prompt to a month of live trading. Including everything that went wrong.

[Leverra @LeverraX]

@stephanlivera Apparently “go all out” belongs in my prompts more often.

[Desimulate @de5imulate]

omw

[Leverra @LeverraX]

@EXM7777 Would love to see a follow-up in a month with the actual traffic results. Especially for a small site starting from zero.

[Leverra @LeverraX]

https://t.co/C1HkGu4U3J

[Leverra @LeverraX]

You can make Claude OR GPT look like the winner without changing a single number.

Just change which row you screenshot.

In Anthropic’s published comparison:

💻 Terminal-Bench 4.0:
Opus 5.5 — 66.4%
GPT-6 Astra — 57.9%

🔬 Terminal-Bench-Science:
GPT-6 Astra — 64.6%
Opus 5.5 — 58.7%

Same table. Two very different headlines.

These are reported scores, not proof of universal superiority. Test settings, safeguards, and uncertainty matter.

But that’s exactly why “the best AI” is an incomplete recommendation.

Best for writing code? Investigating a scientific problem? Finishing your project without hours of corrections?

Before switching models, ask what the winning benchmark actually measures.

Your workload doesn’t have to look like the row that went viral.

[Leverra @LeverraX]

8 agents is interesting. Who can stop all 8 from acting on the same bad signal? The risk controls deserve their own tutorial.

[boan @0xboan]

https://t.co/N5V9XpJlTy

[Leverra @LeverraX]

@xikhar “A few iterations later” is the part that matters. If Opus can keep improving the mechanics without breaking what already works, that’s a serious tool for indie devs.

[Leverra @LeverraX]

@MengTo The atmosphere is what sells this. If Opus can keep this visual consistency as the world gets bigger, small teams could build some seriously good games.

[Leverra @LeverraX]

AI demos need a “second prompt” test.

The first prompt gives you a beautiful website or a playable game.

Then comes the real request:

“Change this one thing. Keep everything else exactly the same.”

That���s the comparison I want to see between Claude and ChatGPT: five rounds of revisions on the same project.

Track what breaks, what stays fixed, and how often a human has to step in.

I’d choose my daily tool based on that.