HawksT

Obsessed with how AI is changing products. Daily notes on what's new and what matters.

[HawksT @John43497285]

@swyx @aiDotEngineer @SydneyScouts When even the people organizing the speakers could probably give a talk themselves lol

[HawksT @John43497285]

@WSJ Going back to look for it and getting caught on camera is quite the ending.

[HawksT @John43497285]

@Reuters Can a company push back on military use and still be considered a reliable supplier? That's a tricky precedent to set.

[HawksT @John43497285]

Curious how much fine-tuning on Singapore-specific scam tactics improves detection over a generic model.

[Google DeepMind @GoogleDeepMind]

Over half of reported crimes in Singapore use sophisticated social engineering.

With @GovTechSG, we’re building on-device fraud detection using our open model Gemma.

By fine-tuning it on local scam tactics, it identifies threats in real time without compromising personal privacy → https://t.co/VdCio2nrHC

[HawksT @John43497285]

OpenAI 这事细节有点离谱。员工把录音设备插在高管会议室电视的 USB 后面，还用卫生纸遮住。更意外的是，公司后来用自家 AI 模型分析了这个设备。

https://t.co/rewR9RNTsZ

[HawksT @John43497285]

@joshwoodward Wait, can you actually tweak the synth controls after it's generated, or do you need another prompt?

[HawksT @John43497285]

@AnnaEconomist Are the capex figures pulled directly from company filings, or are analyst estimates mixed in too?

[HawksT @John43497285]

@Jason Wonder how much of that cost comes from pulling in entire Slack threads every time someone tags it.

[HawksT @John43497285]

4.8x more output tokens and 3.4x more turns at max effort. Curious how much of that extra work actually changes the final code.

[Artificial Analysis @ArtificialAnlys]

Sonnet 5.5 tops the Coding Agent Index at max effort, while xhigh offers an attractive cost-performance trade-off: 63 points versus 68 for less than a quarter of the cost

In Claude Code, @AnthropicAI’s Sonnet 5.5 scores 68 at max versus 63 at xhigh. At $14.19 versus $3.33 per task, choosing xhigh cuts cost by 77% for 5.5 fewer Index points.

Max generates 4.8× as many output tokens and takes 3.4× as many turns. It improves all three benchmark scores, with the largest gain on Terminal-Bench 4.0: 58.1% → 66.2%.

[HawksT @John43497285]

Anthropic 官方出了个 Claude Code 旧代码改造插件 code-modernization。先梳理依赖、提取业务规则，让人确认后再改代码。改完还能用同一批输入对比新旧输出，给每个模块出验证结论。

README 特意提醒，大项目会启动大量 agent，建议先挑一个模块试试。

https://t.co/xdriZrNWHj

[HawksT @John43497285]

@kentcdodds Interesting. Do they ever get stuck going back and forth trying to resolve the same conflict?

[HawksT @John43497285]

@emollick Imagine getting 500 follow-ups about a $12 refund, all perfectly polite.

[HawksT @John43497285]

@BrianRoemmele Does it check what's already in your pantry before adding everything to the shopping list?

[HawksT @John43497285]

@emollick The 3AM staring at the ceiling one is a little too specific.

[HawksT @John43497285]

Curious how Argon handles longer coding tasks where it needs to keep track of changes across multiple files.

[Google @Google]

Gemini 4 Argon began rolling out today. See how devs are already putting Argon to work ↓

[HawksT @John43497285]

Anthropic 的代码日合并量从 2500 涨到 3 万多，Git 已经扛不住了。7 个人花 5 个月造了套版本控制系统 Scion，还顺手做了代码审查工具 Antique。更夸张的是，整个 monorepo 已经迁完了。

https://t.co/RoPN8TtAkZ

[HawksT @John43497285]

@ClementDelangue @huggingface @nvidia Is the bigger bottleneck in France right now GPU availability or getting new datacenters built?

[HawksT @John43497285]

@kentcdodds Curious what happens when two agents touch the same part of the codebase.

[HawksT @John43497285]

323 people making 6,500 edits to a shared company brain is the number that caught my eye. Curious how they keep it from turning into a mess.

[Eric Glyman @eglyman]

https://t.co/SGu5SRww1P

[HawksT @John43497285]

@random_walker The permissions angle is interesting. Two agents running the same model could have completely different risk profiles depending on what they're allowed to touch.