189 wires and counting

$ follow Google

Keep up with Google in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-10-14
stories 96

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

JAX fixes erf_inv/ndtri tails and scatter_max/min gradients

By RepoJournal · Filed · About Google · Composed from the cited sources · methodology

JAX retunes its special-function tail polynomials for accuracy without FMA or float64 and repairs a gradient bug in scatter_max/min with clipped modes, while python-genai fixes transcript and continuity handling across multi-step responses.

Improve tail accuracy of lax.erf_inv, erfcinv, and ndtri without FMA or float64. google/jax

by copybara-service

The Sollya polynomial tables for lax.erf_inv, erfcinv, and ndtri now store only the small remainder around the shifted point, with the dominant sqrt(w) term evaluated in extra precision by splitting into a half-mantissa high part and a residual. Rounding loss in the tail drops across float32 and float64 on CPU, GPU, and TPU without needing hardware FMA.

Propagate scatter mode through scatter_max/min JVP google/jax

by ahlag

The JVP for scatter_max/min only forwarded `mode` to the final scatter_add; internal gathers and scatter_adds ran under the default FILL_OR_DROP. With mode="clip" and out-of-bounds indices the updates were dropped, num_updates went to 0, and the gradient came back inf. `indices_are_sorted` and `unique_indices` now propagate too, so the JVP indexes the same way the primal scatter does.

fix: concatenate output_text across multi-step model responses in a turn googleapis/python-genai

by copybara-service

In the Interactions API, `interaction.output_text` stopped accumulating as soon as its backward traversal hit a `thought` step or a contentless `model_output`, which truncated responses that span consecutive model_output steps. Text is now concatenated across the whole turn, including the interleaved `[thought, model_output, thought, model_output]` shape.

fix(google-cloud-spanner): clear buffered DDL statements on rollback and flush on autocommit googleapis/google-cloud-python

by sakthivelmanii

With autocommit=False, Cursor.execute() buffered DDL in Connection._ddl_statements without setting _spanner_transaction_started, and _reset_post_commit_or_rollback() never cleared that list, so connection.rollback() left unexecuted DDL queued for the next run. Buffered statements are now cleared on rollback and flushed on autocommit.

fix: Match the mTLS endpoint to each transport's SSL context, including plain aiohttp googleapis/python-genai

by copybara-service

mTLS endpoint selection now matches each transport's SSL context, plain aiohttp included, so the endpoint and the context that handles it no longer disagree.

[hijax] avoid using eval_shape in Nonzero google/jax

by copybara-service

Elsewhere in JAX, Nonzero.__init__ computes abstract values and shardings directly instead of going through api.eval_shape, cutting tracing overhead, and the Pallas fuser now treats pl.Blocked(n) as the same block dimension as n (and pl.Squeezed() as None) so fusion no longer depends on which spelling a block shape used. About 20 of the Spanner system tests now share pre-provisioned schemas and...

Quick answers

What shipped in Google on October 14, 2026?
JAX retunes its special-function tail polynomials for accuracy without FMA or float64 and repairs a gradient bug in scatter_max/min with clipped modes, while python-genai fixes transcript and continuity handling across multi-step responses. In total, 45 commits, 50 pull requests, and 1 releases landed.
Who contributed to Google on October 14, 2026?
8 developers shipped this update, including copybara-service, kranipa, ahlag, Yvonne Yu, Chalmer Lowe, renovate-bot, sakthivelmanii, and olavloite.
What were the notable Google updates?
Improve tail accuracy of lax.erf_inv, erfcinv, and ndtri without FMA or float64, Propagate scatter mode through scatter_max/min JVP, and fix: concatenate output_text across multi-step model responses in a turn.