LinkedIn trains AI job search 8x faster

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🔍 LinkedIn trains AI job search 8x faster

🎙️ OpenAI API adds two-way voice

🐋 DeepSeek's cheaper Flash model rivals top rivals

📱 Shopify ditches React Native for native

🐹 TinyGo 0.42 adds recoverable panics

Plus: 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

🔍 LinkedIn trains AI job search 8x faster LINK

LinkedIn detailed the infrastructure behind its AI job search, stacking training optimizations to cut the time needed to distill large teacher models into a compact 0.6B-parameter ranking model by roughly 8x.

Serving teacher models live inside the training loop on SGLang, LinkedIn's online multi-teacher distillation sped work up 3x, while an offline mode precomputes and caches teacher outputs on HDFS or NFS to avoid repeated queries.

LiGer enabled 2x larger batches, with multi-node training, FSDP2, and H200 clusters each adding gains; FP8 was skipped for sub-8B models since casting overhead outweighed savings, and NDCG@10 rose ~24% in production for US users.

🎙️ OpenAI API adds two-way voice LINK

OpenAI opened its GPT-Live-1 speech model as an API, bringing full-duplex voice that listens and talks simultaneously, already powering ChatGPT and letting developers pair it with different backend models to balance reasoning, speed, and cost.

The model runs at $0.05 per minute and posts big benchmark gains over GPT-Realtime-2.1: 80% versus 45% on full-duplex interactivity, tool-calling accuracy up to 87% from 60%, and turn-taking latency dropping to 0.8 seconds.

It ships with twelve new voices across various accents, dialects, and languages, and returns ASR transcripts plus response text by default; Yelp already uses it for phone reservations, reporting better call handling.

🐋 DeepSeek's cheaper Flash model rivals top rivals LINK

DeepSeek today shipped V4.1-Flash, a smaller open-weight model it says beats its larger V4-Pro on quality, speed and cost while matching Claude Opus 5 and GPT-5.6 Sol on coding and terminal benchmarks.

The mixture-of-experts model holds 552 billion parameters but activates only 8 billion while reading a prompt and 16 billion when generating, and a four-bit key-value cache cuts memory to ~890 bytes per token, a quarter of V4-Flash.

Starting September 14 V4-Pro calls reroute to V4.1-Flash at the cheaper rate, dropping peak output pricing from $3.96 to $1.20 per million tokens, roughly 70% cheaper, with weights on Hugging Face under MIT.

📱 Shopify ditches React Native for native LINK

Shopify is abandoning React Native and moving its mobile apps back to separate Swift and Kotlin codebases, six years after adopting the cross-platform framework in 2020 to avoid building every feature twice.

The reversal hinges on AI coding agents now handling enough of the implementation, translation, testing, and review work across two platforms that maintaining parallel native codebases is no longer the deciding cost it was in 2020.

Shopify maintains three React Native libraries: react-native-skia and flash-list are being handed to new maintainers, while restyle, which has a smaller user base, will be archived at the end of 2026.

🐹 TinyGo 0.42 adds recoverable panics LINK

TinyGo 0.42 now supports recoverable runtime panics, letting developers catch nil pointer dereferences, out-of-bounds slice and map accesses, and division-by-zero errors through standard defer and recover, closing a longstanding gap with mainline Go.

Panic handling enables standard library test suites to use Goexit, SkipNow, and FailNow, though fatal conditions like out-of-memory errors stay unrecoverable, so not every runtime failure becomes catchable in constrained targets.

The release adds Go 1.27 and LLVM 22 support plus a new UEFI target that runs Go code natively as a UEFI application before an operating system boots, widening where compiled binaries can execute.

Other news you might like

• Old MacBook uses a mirror, webcam, and AI agent to code its own AMD GPU drivers — 'agent-first' Omarchy Linux debugs itself, AI can check its own progress on screen in real-timeLINK

• Kubernetes v1.37: Scheduler Preemption for In-Place Pod Resize (Alpha)LINK

• Improving Lakebase Postgres compute cacheLINK

• AWS open-sources Pizza Bot: email-style inbox for background AI agentsLINK

• The Pulse: tech companies move to open AI modelsLINK

• Slack can now vibe-code interactive charts and reports inside chatsLINK

• The 60-year-old Unix design choice that made the Linux desktop a messLINK

🧰 Trending tools

Prelint: reviews AI-generated pull requests against your ADRs, docs, and past decisions, flagging product drift before it hits productionLINK

Prefactor: an evaluation layer that scores agent runs in real time, catching quality regressions and drift before they reach production users.LINK

CartAI: automates checkout across merchant sites via one API, handling payments, shipping, and order tracking without building per-site integrations.LINK

ditto: converts any public URL into clean Next.js or Vite code, deterministically extracting components, design tokens, fonts, and hover states via a free API.LINK

Replay QA: records browser sessions with full devtools access and turns them into shareable bug reports, giving AI-built apps QA coverage without writing testsLINK

Browzer: connects your codebase to auto-generate self-healing docs, blogs, quickstarts, and videos, freeing DevRel teams to focus on community and growth.LINK

📚 Trending papers & reports

Monolith-to-microservices splitting automatically carves a tangled legacy app into cleaner standalone services by combining how code connects with how it's named and organized, topping rival tools on three of four tested applications.LINK

Automated bug triage reads incoming software defect reports and routes each to the right handler with about 90% accuracy, cutting the manual sorting that slows down engineering teams.LINK

Code review shortcuts let off-the-shelf language models spot the handful of classes that drive a multi-file code change straight from the commit, beating the best prior tool without any custom setup or training.LINK

AI-plus-human project prioritization gets a governance framework that blends effort estimates 43% more accurately than any single method and flags the shakiest predictions with ~91% detection accuracy, keeping decisions inspectable and recalibratable.LINK

Design conflict detection for connected hardware and software systems predicts when sensor noise or modeling gaps break compatibility, running in microseconds instead of milliseconds and pinpointing the smallest fixes that restore workable designs.LINK

See you tomorrow for a new dose of ☕️ Devshot!