Sep 1 – Oct 1, 2026 · 6 items
The AI week, for a data analyst
A real edition, written for: Data Analyst
The one thing
Which of your weekly analyses could you hand off to a managed agent this quarter?
Simple Jev turns models into decision engines
Featherless AI released Simple Jev. It turns any open model into a classifier or Jev endpoint. The server reads next-token logits to build JSON responses. A public demo API is available with no login or key. It supports choice, score, and noul question types. The demo has a 2k-token context limit and 2 RPS rate limit. Production plans offer higher limits. The evaluation framework includes JevBench scoring and text/decision suites. Local deployment requires Python 3.12+, Transformers, and PyTorch. The server reuses KV cache across question batches. It does not generate or parse JSON completions. The API ignores unknown top-level fields. Chat input uses the model's own chat template.
Why it matters for you
Your SQL and Python analyses often hinge on yes/no or scoring calls. This cuts those to near-zero cost.
- Zero-token decisions for typed choices, scores and yes/no logic
- 82–209× cheaper than frontier models on 2,000-token calls
- Nearly doubled accuracy on real email threads in tests
Try this
Replace one classification step in your next analysis with Simple Jev
Paste this into your AI tool
Context: I analyse business data in SQL and Python Task: classify each row in [paste your dataset snippet] as high/medium/low priority Constraints: - use only the columns shown - return one word per row Output: the priority labels, one per line
Sources
OpenAI DevDay drops new agent tools
OpenAI announced over twenty updates at DevDay 2026. It introduced Dots as always-on agents for ongoing tasks. GPT-6.1 Sol launched with near-Astra performance at one-fifth the token cost. Ultrafast mode offers up to eight times faster token generation in Codex. Private Intelligence now supports zero data retention with private safety processing. Codex can now run in the cloud or via a refreshed CLI with voice control. Code Review in ChatGPT desktop app enables automated and manual reviews. Codex Security Cloud scans repositories and prepares fixes in the cloud. Decisions API entered limited preview for real-time classification. Agents API now supports computer use for software interaction. Bedrock Managed Agents brings OpenAI agents to AWS. Plugin extensions allow sidebar integration in ChatGPT. Improved plugin creation and discovery tools were released.
Why it matters for you
You build analyses that run weekly. Dots can now keep those running without your prompts.
- Dots run 24/7 on tasks like data pulls and report drafts
- Pro/Business Premium plans get first access
- Inspect a dot’s work on its cloud computer
Try this
Ask your admin whether Dots beta is enabled for your plan
Sources
Faster Whisper speeds up audio transcription
SYSTRAN released faster-whisper, a CTranslate2-based reimplementation of OpenAI’s Whisper. It runs up to four times faster for the same accuracy. It supports 8-bit quantization on CPU and GPU. It removes the need for FFmpeg by using PyAV. GPU execution requires NVIDIA cuBLAS and cuDNN. It is installable via PyPI. It supports batched transcription and word-level timestamps. It integrates Silero VAD for speech filtering. It works with Distil-Whisper models. It is open-source under an MIT license.
Why it matters for you
Your Python analyses sometimes start with call recordings. This halves the wait.
- 4× faster than OpenAI Whisper for the same accuracy
- Supports batched transcription and word-level timestamps
- Works on CPU and GPU with 8-bit quantization
Try this
Test faster-whisper on one call recording in your next analysis
Paste this into your AI tool
Task: transcribe this audio file and return the text File: [paste your audio file path] Output: the full transcription
Sources
NVIDIA Rubin smashes agentic inference costs
SemiAnalysis published verified agentic inference results for NVIDIA Rubin. Rubin achieved 67x better throughput per total cost of ownership than GB300. It also delivered 7x better token throughput per megawatt than Blackwell. Rubin earned over 2x more profit per gigawatt than Blackwell. The benchmark used was AgentX, an industry standard for agentic workloads. Results were measured on early pre-release software. Performance is expected to improve as software matures. Rubin outperformed GB300 in P90 interactivity and latency at higher throughput. The upgrade remains justified even at higher rental costs. At 80 TPS, Rubin produced 62% more tokens for the same rental TCO. Across high interactivity scenarios, Rubin achieved up to 16x more tokens per rental TCO.
Why it matters for you
Your weekly analyses could run far more queries at the same compute spend.
- 67× better throughput per total cost of ownership vs GB300
- 7× better token throughput per megawatt than Blackwell
- AgentX benchmark used for real-world agent workloads
Try this
Note Rubin’s numbers for your next compute budget discussion
Agents API offers managed long-running agents
OpenAI launched the Agents API in public beta for all developers. It exposes the same harness and infrastructure that runs Codex. Developers specify a task, model, tools and environment in one API call. Agents can run for hours or days across multiple context windows. The service automatically compresses earlier context as limits approach. Tool search loads only the tool definitions an agent needs. Agents can also split work across subagents that run in parallel. Compute can run in an OpenAI-hosted sandbox, on a developer's own servers, or with partners such as Cloudflare, Vercel, Modal, Oracle and DigitalOcean. The harness is open source, so its logic can be inspected. There are no extra fees beyond tokens and tools used.
Why it matters for you
Your Python scripts that run for hours can now hand off to a managed agent.
- Agents run for hours or days across multiple context windows
- No extra fees beyond tokens and tools used
- Open-source harness means its logic can be inspected
Try this
Check if your API access covers Agents API public beta
Sources
IBM’s Pass^k metric flags flaky agent steps
IBM Research introduced a new metric called Pass^k to measure agent consistency. It found a 24.4-point gap between average success and consistent success for a ReAct agent. The gap arises from flat decision distributions that flip under small perturbations. IBM built a Consistency Analyzer to detect flip-prone steps in agent trajectories. It generates consistency guidelines that reduce the gap from 24.4pp to 12.0pp. The guidelines improve Pass^5 by 16.0pp on same tasks and 13.0pp on similar tasks. Mean accuracy remains unchanged. The method works offline with one recorded trajectory per task. It requires no ground truth or model internals. Full details are in an arXiv technical report.
Why it matters for you
Your analyses fail when one step flips under small changes. This finds those steps.
- 24.4-point gap between average and consistent success in tests
- Consistency Analyzer detects flip-prone steps in recorded trajectories
- Guidelines cut the gap from 24.4pp to 12.0pp
Try this
Record one agent trajectory and run IBM’s Consistency Analyzer on it
Paste this into your AI tool
Task: analyse this agent trajectory for flip-prone steps Trajectory: [paste your recorded agent steps] Output: list of steps most likely to flip under small changes
Build this
Every week, one small thing to build with AI in something you actually care about. No work in it. Five minutes to set up, and worth keeping if it earns a second run.
A weekly plan that only uses the analyses you already run
Five minutes to set up
Context: I am a Data Analyst building weekly dashboards and analyses Task: build a weekly plan using only the analyses I already run Inventory: [list the analyses you already run] Output: the weekly plan with each analysis scheduled
- List every analysis you already run weekly
- Use ChatGPT to order them by dependency and effort
- Check the plan covers all your current outputs
- Save the plan as your new weekly template
Yours arrives Thursday.
This one was written for a data analyst. Tell us what you do and the next one is written for you — same news, your job, once a week.