Anthropic's own data: junior hiring fell from 27% to 9%

Good morning 👋 Anthropic just handed out the receipts on its own junior-hiring collapse, the kind of number most companies bury in a trend piece instead of publishing themselves.

In today's issue:

• 🔭 Anthropic's own data: junior technical hiring share fell 27%→9%
• 🧠 Black Forest Labs ships a 6B open image model that renders in under 2 seconds
• 🔬 An agent benchmark that humbles every model, LLM-designed quantum circuits, and a fix for lazy MoE routing
• 🛠️ Turn a document pile into a slide deck, minus the hallucinated numbers
• 🔗 Anthropic's own expectations gap, NVIDIA's city dispatchers, and three more

🔭 THE ONE THING

📉 Anthropic's own data: junior hiring fell from 27% to 9%

Anthropic published its own employment data this week, and the hiring line is the one that matters: junior technical hires fell from 27% of new hires in early 2025 to 9% by mid-2026, and junior headcount in technical departments slid from 21% to 14% over the same stretch. By August, interactive agent hours outpaced employee hours 6 to 1, combined agent hours passed human hours company-wide back in March, and engineers' commit velocity is running 5x where it was six months ago. One in five staff told the company in April that their job had turned into managing Claudes rather than doing the work. I buy the productivity claims: a company pulling 2 to 4x the revenue per employee of Meta or Google at the same growth stage isn't padding numbers for a press release. The agent stats are the flashy number; the one to sit with is that the company building the model already has receipts showing its own entry-level hiring shrinking.

🧠 MODELS & RELEASES

• 🎨 Black Forest Labs shipped FLUX.3 Image Klein, a 6B open-weight image model, in two flavors: a distilled checkpoint and a base one. The distilled version renders a 1024×1024 image in 1.93 seconds on an H100, in 16 steps instead of the base model's 50. Hugging Face has DreamBooth LoRA scripts and quantized (fp8, nvfp4) builds ready day one, so fine-tuning and running it on smaller cards isn't a side project. Welcome FLUX.3 Image Klein in Diffusers

🔬 RESEARCH HIGHLIGHTS

• FDE-Gym — A new benchmark drops agents into the actual job of a forward-deployed engineer, not a clean ticket. 95 tasks, 20 customer scenarios, 18 sectors, built from real delivery work. The best of 13 model setups across seven providers scored 0.56 out of 1. Only 14 of 95 tasks cleared the bar for "solved." Hand a model every missing fact from a cooperative stakeholder and it still only clears that bar on 10 of 23. The bottleneck isn't reasoning. It's figuring out what the client actually meant. arXiv

• Quantum circuit discovery — An LLM proposes circuit designs, an independent evaluator scores and verifies each one, and the loop runs until something wins. Across eight tasks spanning chemistry, optimization, and number theory, the discovered circuits beat the hand-tuned state of the art on seven of them, cutting Toffoli gate and qubit counts. On the eighth, the system found a new architecture for modular inversion that makes a previously impractical classical algorithm usable. Narrow, task-specific search, not a general quantum-algorithm designer. Still a real engineering result, not a leaderboard score. arXiv

• MirrorMoE — Takes aim at the fine-grained-experts-plus-shared-experts combo most frontier MoE models now run. The authors' case: shared experts flatten token representations enough to choke the extra routing exploration that fine-graining is supposed to unlock. Net effect, a chunk of that extra expressiveness never gets used. Fix: suppress experts that aren't earning their keep for a given token. Reportedly lifts performance across model scales, no hard numbers in the abstract. Niche, but it explains why stacking more experts stopped paying off for some teams. arXiv

🛠️ TRY THIS

Hand the deck-building grunt work to a RAG pipeline instead of a junior hire

1. Drop your source docs (specs, research notes, last quarter's deck) into a bucket an agent can retrieve from.

2. Split the job across three agents: one plans the outline, one pulls grounded facts via RAG, one drafts the slides. AWS's Bedrock AgentCore walkthrough is the cleanest version of this pattern I've seen.

3. Put guardrails in front of anything customer-facing. A hallucinated number in a board deck costs you more than a slow analyst ever did.

4. Time it end to end. AWS clocked theirs under 90 seconds. Slower than that, and the bottleneck is retrieval, not the model.

Prompt: Using only the attached documents as source material, draft a 10-slide outline with one supporting fact per slide and a citation back to the source doc. Flag anything you can't ground.

🔗 QUICK LINKS

• Tamnoon built an agentic remediation layer on Bedrock that cut client cloud exposure more than 90% in 90 days, with zero production incidents. AWS writeup
• NVIDIA is powering physical AI agents that EMEA cities are now using to run traffic, energy and maintenance calls. Less demo, more dispatcher. NVIDIA blog
• Stack Overflow shipped an agent marketplace in a few weeks using Codex, and the ChatGPT plugin alone drove up to 10x more agent creation. OpenAI
• Anthropic's own staff expect AI to automate 75% of their work by 2030. The US public's guess: 28%. One of those groups is pricing the next four years wrong. research
• Anthropic is borrowing like it believes that forecast: Blackstone is anchoring $9B of a $60B TPU financing package at a 10%+ yield, per the FT. FT

Three separate Anthropic stories in one issue today. Didn't plan it that way, but there it is.

Pradeep Perugu