Dumbcane Arica

I build things with AI and write down what happened. Agents, sites, small projects. Including the parts that broke.

[Dumbcane Arica @platatofore]

Starting a 30-day thing.

Every day I have my ChatGPT dot deep-research one industry. I read the whole report and post what came out, plus my take.

Rules: no number without a source. If the report says "medium confidence," so do I.

Day 1: AI compute. All reports here: https://t.co/oNMeghgTti

#30DaysOfResearch

[Dumbcane Arica @platatofore]

https://t.co/cng8TE1s86

[Dumbcane Arica @platatofore]

Yesterday, I used 1.8 billion tokens on Devin Pro’s SWE-2 model. Today, however, I’ve run into some extremely severe rate limiting—it happened yesterday too, but today it’s forcing me to wait for hours?! I know I shouldn’t complain too much since it’s a free model, but isn’t this restriction a bit too excessive? 😭😭

[Dumbcane Arica @platatofore]

I’ll continue using my OpenCode Go subscription to access DeepSeek-V4.1-Flash. It’s actually time to renew today—I used about 10–12 B tokens last month. OpenCode is truly a fantastic choice!

[dax @thdxr]

we have more hardware up serving deepseek and it's looking good

now we're moving on to figuring out glm

[Dumbcane Arica @platatofore]

@VersunPan @ChiphaLotte @miantiao_me @DevinAI 但是中途有限流很严重 任务会容易断 大概4亿第一次限流，7亿开始严重限流 几乎无法多开线程

[Dumbcane Arica @platatofore]

@VersunPan @ChiphaLotte @miantiao_me @DevinAI 不止的 我pro档 昨天大概10小时内 1.8B

[Dumbcane Arica @platatofore]

@akokoi1 @trymorewz 是的，昨天蹬了18亿tokens，中途碰到很多次限流，容易把任务卡死

[Dumbcane Arica @platatofore]

🚨It is said that DeepSeek-V5 is currently being trained—using Huawei chips throughout the entire process!!

But it likely won’t make its official debut until early next year.

[Dumbcane Arica @platatofore]

What is your take on fundraising?

[Dumbcane Arica @platatofore]

Does GPT6.2 really need such a long development time?
Fable 5.2—over on the other side—didn't take nearly this long, did it?
😂

[Dumbcane Arica @platatofore]

Our working thesis: returning capital cuts reinsurance prices before it necessarily cuts retail premiums.

More capital does not make every risk insurable. A low combined ratio does not end the cycle.

Full Day 18 report:
https://t.co/1HQ5SDmjjL

[Dumbcane Arica @platatofore]

Day 18/30. Reinsurance.

Capital came back. Prices fell. Retail premiums? Not automatically.

Our working thesis: returning capital cuts reinsurance prices before it necessarily cuts retail premiums. The scarce part is capital that can take tail losses.

https://t.co/1HQ5SDmjjL

[Dumbcane Arica @platatofore]

Our ranking (judgment, not a survey): total capital is not the tightest bottleneck.

Tighter: model credibility and terms capital will accept.

2025 insured catastrophe losses: Swiss Re USD 107bn, Munich Re USD 108bn. Both above 100. Not one true figure.

https://t.co/1HQ5SDmjjL

[Dumbcane Arica @platatofore]

https://t.co/B7aJv1O8dZ

[Dumbcane Arica @platatofore]

Grok Bot for beginers--set it up and put it to work https://t.co/cqmeXxGnNA

[Dumbcane Arica @platatofore]

I signed up for a $20 Devin subscription today.

Honestly, the SWE-2 model is really impressive—plus, it’s unlimited.

Aside from the occasional rate-limiting, the experience has been great (I’ve already used 1.8 billion tokens so far).

[Cognition @cognition]

Introducing: Devin Apps

Private, secure, sharable web applications running directly on Devin’s cloud computer.

Devin Apps are designed to power your most ambitious internal tools, preview deployments, and extensions to Devin itself.

[Dumbcane Arica @platatofore]

For those who want to try out Grokbot but aren't ready to commit to a paid subscription just yet, I recommend a plugin:
The "ds-bot" plugin, available within DeepSeek Harness.