Ambitious AI

The sports channel for people who are ambitious. Scores, highlights and hot takes every day. Join the free app 👇

[Ambitious AI @Ai_ambitious_]

COGNITION 🔥: SWE-2 Flash is in Devin Desktop and CLI.

- mid-trained on Inkling-Small (276B total, 12B active)
- one-fifth the price and 3x the speed of SWE-2
- 200 tok/s vs SWE-2's 60

Weights not public yet. Partner program coming.

— Cognition

[Ambitious AI @Ai_ambitious_]

@levelsio ported Battlefield 1942 to mobile web with Claude Code.

> built a graphics layer that translates Direct3D to WebGL
> sits between the game's machine code and the browser
> fixed portrait HUD: perspective draws vs flat HUD draws

Full multiplayer in the phone browser.

[Ambitious AI @Ai_ambitious_]

@ladyxtel's Living Brain: one memory across ChatGPT, Claude, Cursor, and Codex.

> add their MCP connector
> sign in with your Living Brain account
> write in one tool, retrieve in any

92.2% on LoCoMo. 464ms median retrieval. Up to 22.6x fewer tokens per question.

[Ambitious AI @Ai_ambitious_]

API tip for Images 2.6 edits:

> start most edits on Flare at low complexity
> raise complexity only when the scene has many constraints
> set image quality separately
> mark up the image to show exactly where to change

Model id: gpt-image-2.6-flare

Per OpenAI Devs

[Ambitious AI @Ai_ambitious_]

OPENAI 🔥: Images 2.6 is live in ChatGPT, Codex, and the API.

- generates images 50% faster
- GIFs from a prompt or a still
- Flare for speed, Sunburst for detail

Flare at low quality: as little as 6 seconds.

— OpenAI

[Ambitious AI @Ai_ambitious_]

Tip via @levelsio: have Claude Code audit your infra bills.

It found him ~$34K/mo in savings. One easy fix: his backups kept every file version for a decade. A 30-day lifecycle rule took Backblaze from $1,200/mo to $26.

Check your bucket's versioning settings tonight.

[Ambitious AI @Ai_ambitious_]

@xikhar built a playable 3D game in 24 hours:

> Flux 3 Max makes a reference video to lock the art style
> rebuild its scenes in Blender, place them in Three.js
> GPT Image + Tripo turn the hero into a 3D model
> Claude Code agents build each level

Play, report bugs, repeat.

[Ambitious AI @Ai_ambitious_]

Revolut made a self-hosted open model the default for ~2,000 engineers.

- GLM-5.3 on 56 Nvidia B200s
- ~30 cents per 1M tokens, about a tenth of its old API bill
- engineers stopped rationing tokens and handed off Jira, docs and debugging too

— Revolut's David Abitbol

[Ambitious AI @Ai_ambitious_]

@levelsio's whole vlog edit prompt:

"Claude can you edit these clips into a 10 minute vlog for X and YouTube"

Run Claude Code in the clips folder. Opus 5.5 transcribed them with mlx-whisper, made frame grids to see each clip, wrote the cuts as JSON.

58 min in, ~10 out.

[Ambitious AI @Ai_ambitious_]

TENCENT 🔥: Hy4 is out, and it can see now.

- first Hunyuan model with vision and computer use, CLI and GUI, trained on Blender and CAD
- 1M-token context
- $0.834 in, $2.501 out per 1M tokens, half off through Dec 31

On OpenRouter as tencent/hy4. Open weights promised soon.

[Ambitious AI @Ai_ambitious_]

Honest question for people running coding agents all day

How do you know a task is actually done?

> You read the diff
> Tests pass
> A second agent reviews it
> You ship and find out

I think most agent bugs slip through because "done" was never written down

[Ambitious AI @Ai_ambitious_]

Anthropic's Residency is open, and you don't need an ML background.

- for technical founders, staff engineers and quant professors
- 6 weeks, 3 months or 6 months
- pays $350K to $500K a year equivalent, prorated
- remote possible, SF preferred

Apply by Nov 22.

— Anthropic

[Ambitious AI @Ai_ambitious_]

Claude Code prompt for briefs with no fake sources:

/goal write a one page brief on [topic] where every claim has at least three sources and every link opens to a real page that backs it. Open each link. Replace any that don't.

Claude keeps going until it passes.

Via @Raytar

[Ambitious AI @Ai_ambitious_]

ANTHROPIC 🔥: Claude's API has message threads, in public beta.

- send the conversation once, then only new messages
- median time to first token down ~35% on 500K-1M token Claude Code requests, per Anthropic
- same prices and rate limits

Threads expire after 5 idle minutes.

[Ambitious AI @Ai_ambitious_]

Mistral had its open-weight Large 4 build a live 3D tracker of the ~19,000 satellites and debris pieces orbiting Earth. It shipped a working app.

Try it:
> model id: mistral-large-4
> $0.70 in, $2.10 out per 1M tokens
> 1M-token context
> OpenRouter, or self-host with vLLM

[Ambitious AI @Ai_ambitious_]

ANTHROPIC 🔥: Claude's $200 plan no longer has a 5-hour limit for verified accounts.

- Max 5x is now "Max" ($100), Max 20x is now "Ultra" ($200)
- Most existing subscribers have it now; newer accounts wait for verification
- Weekly limits unchanged

— Anthropic

[Ambitious AI @Ai_ambitious_]

@theo Source:

[Ambitious AI @Ai_ambitious_]

opus 5.5 tip:

drop reasoning effort from Extra High to High.

@theo switched a week or two ago and says it's "better almost all of the time."

keep Extra High for the long, hard jobs. use High for everything else.

[Ambitious AI @Ai_ambitious_]

Flipkart’s quick-commerce unit hits 1.1 million to 1.2 million daily orders—nearly triple its November 2024 volume—two years after launch.
Can a quick-commerce player scale this fast without deep-pocketed logistics backing?

[Ambitious AI @Ai_ambitious_]

Mark Walter’s sports empire runs on insurance-backed financing—where policies on future revenue lock in credit lines. That’s not just leverage; it’s a closed-loop capital engine built on risk transfer.