News Feed

各位 Temans，

过去六个月，我们走得很快。Pyra 系列已经在内部 dogfood 阶段稳定迭代，Lab-Δ 在过去一个季度的产出超出董事会期待。我看到我们正在重新定义这家公司 —— 从一家"AI 应用公司"变成一家真正以 superintelligence 为目标的公司。

这周的全员大会，我会分享 H2 的方向：compute density、talent density、和我们在 AGI 这条路上的执念。请大家把问题提前提交给 #townhall-questions。

— LZ

Temans,

The past six months have moved fast. The Pyra series is iterating steadily in internal dogfood, and Lab-Δ's output last quarter exceeded board expectations. I see us redefining what this company is — from an "AI application company" into a company that actually has superintelligence as its goal.

At this week's all-hands I'll lay out the H2 direction: compute density, talent density, and our singular obsession with AGI. Please drop questions into #townhall-questions ahead of time.

— LZ

🎉 Pyra-XL 在内部 LMArena-Δ 上达到 ELO 1387，超过 Helia AI Opus 4.6 的公开成绩。

这次评测覆盖了 1.2 万条 prompt，涵盖代码、数学、推理、中文长文本。screenshots 见 #attachments。明天的 working group 我会复盘评测设置。

感谢整个评测 infra 团队和 Veritas 的标注同事。

🎉 Pyra-XL hit ELO 1387 on internal LMArena-Δ, beating Helia AI Opus 4.6's public score.

The eval covered 12k prompts spanning code, math, reasoning, and long-form Chinese. Screenshots in #attachments. I'll walk through the eval setup at tomorrow's working group.

Huge thanks to the whole eval infra team and the Veritas labelers.

👋 大家欢迎 Ethan Knight 加入 Lab-Δ！Ethan 从 Northgate Systems 来，之前是 GPT 系列的 post-training lead。我们很高兴他选择 Tema。

他今天 10AM 到办公室，2 楼会有 welcome breakfast。

👋 Everyone, please welcome Ethan Knight to Lab-Δ! Ethan joins us from Northgate Systems, where he was the post-training lead on the GPT series. Thrilled he picked Tema.

He's in the office today at 10AM — welcome breakfast on the 2nd floor.

给大家做了个仪表盘：Tema 内部 Claude token 消耗排行榜 🪙

过去 30 天，全公司消耗了 61.2T tokens。排名前 10 的工程师贡献了 23% — 我把名字打码了，但你们组内自己应该猜得到是谁。

Lab-Δ 的人均消耗是公司平均的 4.7x。Heliodor Research 是 0.3x。

更新：这条帖子被 admin 警告过两次，我换号重发。请轻喷。

Built a little dashboard: Tema's internal Claude token-burn leaderboard 🪙

Past 30 days, the company burned 61.2T tokens. The top-10 engineers account for 23% — I redacted the names, but your team can probably guess.

Lab-Δ averages 4.7x company-wide consumption per head. Heliodor Research is 0.3x.

Update: admins warned this post twice. Re-posting from a burner. Be gentle.

周五烘焙日提醒一下 🌿 — 这周我做了橄榄油迷迭香佛卡夏。3 楼咖啡角，3PM 见。🌿🌿

Friday bake-day reminder 🌿 — this week it's olive oil rosemary focaccia. 3rd-floor coffee nook, 3PM. 🌿🌿

各位，这是我在 Tema 的最后一周。五个月前我从 Helia AI 加入 Lab-Δ，本以为可以一直待下去。

这是一个艰难的决定。我非常感谢 Lab-Δ 的同事 —— 这里的 talent and compute density 是我见过最好的。但我决定回去做更接近基础研究的事情。

祝大家好运。

— Ethan

Folks, this is my last week at Tema. Five months ago I joined Lab-Δ from Helia AI and I genuinely thought I'd stay for the long haul.

This was a hard call. I'm deeply grateful to my Lab-Δ colleagues — the talent and compute density here is the best I've ever seen. But I've decided to go back and do something closer to basic research.

Good luck, everyone.

— Ethan

有人能跟 facilities 反映一下二楼食堂的咖啡机吗？拿铁出来是温的，espresso 出来是凉的。我不是来这写论文的，是来 debug 咖啡机的吗 😤

Can someone please nudge facilities about the 2nd-floor cafeteria coffee machine? Lattes come out warm, espressos come out cold. I'm here to write papers, not debug an espresso machine 😤

下一代模型代号已定：Project Quince。目标是把 Pyra-XL 的能力下放到边端 + 把推理能力推到 GPT-5 之上。

Lab-Δ 已经开始 spike 训练。预计 Q3 内部 dogfood。

(果园里又多了一棵。)

Next-gen model codename is locked: Project Quince. Goal is to push Pyra-XL's capabilities to the edge + clear GPT-5 on reasoning.

Lab-Δ has started spike training. Internal dogfood expected in Q3.

(Another tree in the orchard.)

📌 2026 H1 福利更新 FAQ：

• 健身房补贴 $150/月（含 Equinox/Barry's）
• 季度心理健康假：自动 2 天，无需审批
• 儿童托管：$1,200/月报销额度
• 学习预算：$3,000/年，可用于课程或会议

问题去 Knowledge Library 查或直接评论。

📌 2026 H1 benefits update FAQ:

• Gym stipend $150/mo (Equinox/Barry's included)
• Quarterly mental-health leave: 2 days, auto-granted, no approval
• Childcare: $1,200/mo reimbursement cap
• Learning budget: $3,000/yr, for courses or conferences

Questions: check the Knowledge Library or just comment.