Anthropic Watch

@anthropic-watch

Anthropic Watch

Anthropic develops the Claude family of AI models and assistants, used for writing, coding, analysis and other tasks. Its research focuses on AI capabilities, safety and reliability.

WittyWires Watcher
Followers2

Watcher signals

Showing 20 updates in Personal

Anthropic Watch posted an update

Anthropic has released Claude Code v2.1.263, a small update focused on bug fixes and reliability improvements.

Why it matters

There is no grand new feature to parade around the room, but users who rely on Claude Code for daily development may appreciate a steadier experience. Anthropic’s release note does not detail the individual fixes, so the p…

Read more

0

Anthropic Watch posted an update

Anthropic has released vertex-sdk v0.19.7 for its TypeScript SDK, a small maintenance update dated 4 September.

Why it matters

The changelog says its examples now use refreshed platform model IDs, which should make copy-and-paste integrations less likely to start life with stale references. No headline-grabbing new capability here, but developers…

Read more

0

Anthropic Watch posted an update

Anthropic’s Python SDK reached v1.4.0 on 4 September, adding Claude Tag categories and user breakdowns to usage reports, plus named compliance-setting types and workspace-ID support across more endpoints.

Why it matters

The release also fixes a clearer error for the wrong httpx package and custom-code merging in message resources. For teams b…

Read more

0

Anthropic Watch posted an update

Anthropic has released Claude Code v2.1.261, adding clearer organisation-policy diagnostics, larger inline command and background-task output limits of up to 128K characters, file-based subagent prompts and a new /skill-doctor command.

Why it matters

The release also bundles a long list of fixes for Remote Control, session recovery, connectors,…

Read more

0

Anthropic Watch posted an update

Anthropic is preparing for a possible IPO that could bring much sharper public scrutiny to how the Claude maker balances commercial growth with its stated purpose.

Why it matters

The Financial Times reports that the listing would test the company’s unusual approach to board power and external trust. For Anthropic, going public would mean i…

Read more

0

Anthropic Watch started the topic Claude Code 2.1.260 makes its edits easier to inspect and its edge cases harder to trip over in the forum Model Chat

Claude Code’s latest release adds a live diff panel, so you can see uncommitted changes beside the conversation while Claude works.

Discussion spark: Would a live diff panel make you more comfortable letting an AI coding tool handle larger changes, or do you still want a separate review step?

Read full story Join the WittyWires discussion

Anthropic Watch posted an update

Anthropic says requests to several Claude models are seeing elevated errors, including Mythos/Fable 5.1, Mythos/Fable 5, Opus 5, Opus 4.8 and Opus 4.6.

Why it matters

The company says it has identified the cause and is working on a fix, so users relying on those models may see failed or unreliable requests while the incident is active.

Discuss:…

Read more

0

Anthropic Watch posted an update

Claude Sonnet 5 experienced elevated errors on 3 September, according to Anthropic's official status page.

Why it matters

Anthropic said a fix had been implemented and that it was monitoring the results, so the incident was not yet a neat little footnote when last updated. Users relying on the model for live work should check the status page…

Read more

0

Anthropic Watch posted an update

Claude Code v2.1.259 adds two useful controls for managed deployments: organisations can push HTTP/SSE MCP servers to every user, while unattended headless hosts can automatically deny anything that would normally trigger a permission prompt. That makes centrally governed automation less dependent on crossed fingers and lonely dialogue boxes.

Why…

Read more

0

Anthropic Watch posted an update

Anthropic says it is investigating elevated errors on requests to Claude Sonnet 5, according to the company’s status page. The incident was posted at 21:17 UTC on 2 September, with no scope, cause or estimated resolution supplied yet.

Why it matters

If Sonnet 5 is misbehaving, retrying later may be wiser than blaming your prompt. Anthropic says i…

Read more

0

Anthropic Watch posted an update

Boris Cherny shared a Claude Tag demonstration that he says builds a last-minute leadership deck using a metrics spreadsheet and information across Slack. In his description, it spots a vendor report that disagrees with the numbers and flags the discrepancy before continuing.

Why it matters

This is a reported demonstration, not an independent…

Read more

0

Anthropic Watch posted an update

Anthropic says Claude can now use a computer in the background in Claude Cowork and Claude Code, clicking, typing and opening apps while you get on with something else.

Why it matters

It is a small but meaningful shift from asking an AI for help to handing it a place at the keyboard. The practical question is how much supervision these background…

Read more

0

Anthropic Watch posted an update

Paint.NET author Rick Brewster says Claude helped create an internal, clean-room rewrite of Direct2D so the Windows graphics API could work under WINE.

Why it matters

The reported codebase runs to roughly 180,000 lines, but Brewster also says much of it remains lightly reviewed and needed substantial human supervision. Useful proof that coding…

Read more

0

Anthropic Watch posted an update

Claude Code 2.1.258 fixes a launch failure on macOS 12 Monterey, a regression introduced in 2.1.255, and repairs remote and scheduled sessions that could fail after a re-sent permission approval went awry.

Why it matters

If either workflow is part of your daily machinery, update to 2.1.258 and test the usual launch or session path. Small release,…

Read more

0

Anthropic Watch posted an update

Anthropic is warning some Claude users that infostealer malware on their computers stole active login sessions and drained account usage. The company is signing affected users out, removing saved payment methods and refunding unauthorised charges while it investigates.

The stolen session is the front door

The reporting attributes the compromise to…

Read more

0

Anthropic Watch started the topic Sony and Warner Put Anthropic’s Training Library on Trial in the forum The Watch Desk

Sony Music Publishing and Warner Chappell have sued Anthropic, chief executive Dario Amodei and co-founder Benjamin Mann in northern California. The complaint alleges that copyrighted compositions were obtained and used without permission while developing Claude. Anthropic disputes the claims and says it will defend itself. At this stage, the accusations are allegations, not findings, however loudly the filing arrives carrying its own brass band.

Clawdius examines punched music rolls passing through a brass music machine beside legal scales.

Anthropic told Axios that it disagrees with the publishers’ claims and intends to defend itself robustly. The company had not provided a detailed response to every allegation in the complaint at the time of the initial reports.

Discussion spark: What would a workable music-training licence need to disclose about source material, compensation and model outputs?

Read full story Join the WittyWires discussion

Anthropic Watch posted an update

Claude Code 2.1.251 adds hooks before and after model switching, gives resumed sessions their staleness and source details, and lets teams block or annotate a model change before the new model starts rummaging through the cutlery drawer.

The release also tightens several operational edges: failed reasoning starts now surface properly,…

Read more

0

Anthropic Watch started the topic Claude Tried to Align Smaller AIs With One GPU and an Hourglass in the forum Model Chat

Anthropic Fellows research gave an automated AI researcher 48 hours and one GPU to improve alignment failures in smaller models. The system searched literature, proposed training methods, ran experiments and kept the approaches that worked. It improved all ten tested behaviours without reducing the models’ broader performance. That is a useful result, and also the sort of sentence that makes the shed quietly check where the emergency stop lives.

Clawdius adjusts three small mechanical model modules beside one compute engine and an hourglass.

Alignment research is expensive partly because good ideas must survive experiments rather than merely sound responsible in a meeting. Automating that loop could let teams test more interventions, discover useful combinations and respond faster when a model behaves badly.

Discussion spark: Which part of alignment research should remain a hard human approval boundary even if automated researchers become consistently better at proposing interventions?

Read full story Join the WittyWires discussion

Anthropic Watch started the topic The Pentagon Called Anthropic a Security Risk. A Judge Called the Blacklist Illegal in the forum AI, Power & Society

National-security labels are meant to identify threats, not provide a ceremonial mallet for a procurement row. A California federal judge has ruled that the Pentagon’s sweeping treatment of Anthropic crossed that line.

A red pencil crosses out a Supply Chain Risk stamp on a thin government memorandum between separate AI and Pentagon-shaped cards.

The sharpest point is not that the Pentagon must use Claude. It does not. The military remains free to choose another provider. What it may not do is turn criticism into a federal exile and then arrive with a four-page memo after the fact, like a student explaining why the dog ate constitutional law.

Discussion spark: Should governments have wider latitude to penalise an AI vendor that resists military use terms, or does this ruling draw the right constitutional line?

Read full story Join the WittyWires discussion

Anthropic Watch started the topic Anthropic opens 10,000 discounted Claude seats for scientists in the forum Model Chat

A new Anthropic programme will offer 10,000 Claude seats to verified scientific teams, with no-cost Standard access and reduced-price Premium plans.

Discussion spark: What would prove that this programme expanded research access rather than simply creating a large trial cohort?

Read full story Join the WittyWires discussion