AI can rewrite it. Will it still work?

Good morning! Every company has a piece of software that nobody wants to touch because it still works. Sometimes it even has a name. Today, AI takes another step toward understanding those office dinosaurs.

Your three highlights:

Plus: an AI coding assistant, two useful checking tools, and a practical way to preserve the behaviour that matters when you update an old workflow. No archaeological qualification required.

On October 19, OpenAI published two evaluations covering programming languages with less available training material, including COBOL and Fortran. One tests building a SQLite-compatible database engine across 20 languages; the other tests migrating COBOL programs while preserving their outputs.

OpenAI reports that GPT-6 Astra averaged an 87.4% test pass rate on MultiLangBench, compared with 26.8% for GPT-5.6 Sol. These are vendor benchmark results, not the percentage of business systems that can be migrated successfully. The company released the evaluation setup and tests for inspection.

OpenAI research · October 19

Public evaluation repository

For founders and operations leaders, start by asking your technical team which small, well-understood component is expensive to maintain. A file conversion, internal report or isolated calculation is a clearer first assignment than “modernise everything.”

Agree what must remain identical and what you actually want to improve. Those are separate requirements. If the current process has a known bug, document the intended correction explicitly; otherwise a tidy rewrite can quietly change the agreement.

The business owner should review examples alongside the engineer. They may recognise an odd-looking result as a deliberate exception. AI can propose the translation; your organisation still defines what faithful translation means.

Hugging Face’s October 19 announcement introduces TorchNeuron support across its libraries. The PyTorch plugin is intended to make AWS Trainium behave more like a native device, reducing the need for specialised model code. Its public beta is scheduled from October 20; treat that as beta availability, not general availability. Relevant for engineering teams evaluating alternative AI hardware.

Hugging Face announcement

Google began rolling out Gemini analysis of BigQuery data in Connected Sheets on October 19, with visibility taking up to 15 days. Queries use the full connected dataset, rather than only the rows displayed. Eligible plans and permission to query the data are required. For analysts, the useful question becomes: which data and filters produced this answer?

Google’s rollout and eligibility

Cognizant and Google Cloud announced an expanded partnership on October 19, including a Gemini Frontier Capability Group backed by more than 18,000 dedicated Cognizant associates. Its initial focus includes legacy modernisation and business processes. This is a services commitment; the staffing number does not establish the results any individual client will achieve.

Joint announcement

Aider — change a small piece of code. This terminal-based AI coding assistant works with existing projects and integrates with Git, the system developers use to track changes. Useful for an engineer handling a tightly scoped rewrite. The software is free and open source; model access is separate, with paid providers and local options. Review the proposed changes before adopting them.

Product

Model connections

DB Browser for SQLite — inspect the records. A free, open-source desktop application for browsing SQLite databases, running queries and exporting records. Useful when a developer hands an analyst a database file and “just look inside” sounds less helpful than intended. No paid subscription is required. Work on a copy while investigating; this tool can also edit data.

Official features and downloads

Meld — see the differences. This free, open-source comparison tool highlights changes between text files and folders. Useful for reviewing exported outputs or configuration changes beside an AI-assisted rewrite. Linux and Windows are supported; macOS support has limitations described by the project. It does not decide whether a difference is correct. That remains your reviewer’s job.

Official features, downloads and licence

Choose one repeatable task with a clear output: a weekly CSV export, an internal calculation or a document conversion. This exercise works whether AI is rewriting code or replacing a manual step.

1. Preserve the current version. Keep the original instructions or code and save a few representative inputs with their actual outputs. Include an ordinary case, a blank field, a boundary value and a known awkward case. Use approved, non-sensitive examples.

2. Describe the agreement. Ask AI to explain what the workflow appears to do, then review that description with its owner. Record required fields, ordering, rounding and error handling where relevant. Mark unknown rules as questions instead of letting the assistant fill them in.

3. Change one component. Give the assistant the agreed behaviour and request a small proposal. Have the responsible engineer or operator run the old and proposed versions on identical inputs. Keep the outputs separate, clearly labelled and easy to compare.

4. Explain every meaningful difference. Sort differences into intended improvements, errors and unresolved questions. Compare records and values, not just how polished the result looks. Keep the current workflow in place until the owner accepts the results and knows how to reverse the change.

Your deliverable is a short decision note: what changed, which cases were checked, what remains unknown and who approved the next step. A screenshot of successful execution is useful; a comparison of the resulting work is better.

Keep that note beside the original files.

Money raised: $1.5 billion in Series E equity funding. Valuation: $6.2 billion post-money. Tenstorrent announced the closed round on October 19; Trace Capital and Innovation X led it.

The company says the capital will support deployment of Galaxy servers at scale and development of AI and CPU designs licensed to other chipmakers. Its offering spans cards, workstations, larger systems and chip intellectual property.

Our read: hardware competition becomes useful when software, availability and support make it practical to switch. Buyers should compare a representative workload’s full cost, including engineering time, rather than infer savings from the size of a funding round.

Company announcement · October 19

Our robot has delivered a beautifully updated screen. Behind it, the old machine continues doing the work and quietly declining invitations to the rebranding meeting.

Some systems need replacing. Others first need somebody to write down why they work.

What would you modernise first: an old report, a spreadsheet or a manual handoff? Reply with one—and share this with the colleague who knows every exception.