diff --git a/archive.html b/archive.html index b2746a7..c11075a 100644 --- a/archive.html +++ b/archive.html @@ -115,9 +115,14 @@

All digests

-

98 issues

+

99 issues

- Sunday, August 2 2026 + Monday, August 3 2026 @@ -431,33 +431,29 @@
Today's Digest
- Sunday, August 2 2026 + Monday, August 3 2026
-

Anthropic's new top-tier model gets close to its most capable one at half the price — the clearest sign this week that the expensive end of AI keeps getting cheaper.

+

Alibaba has a new top-tier coding model, Californians can now force data brokers to delete their records, and someone squeezed a language model onto a 1970s home-computer chip.

Lead Story
-

Anthropic Releases Claude Opus 5

-

Anthropic released Claude Opus 5, the newest model in its Opus tier. It is aimed at tasks that run for a long time without someone watching over them, and Anthropic says it comes close to the intelligence of its most capable model, Claude Fable 5, at half the price.

+

Alibaba's Qwen Team Releases Qwen3.8-Max

+

Alibaba's Qwen team released Qwen3.8-Max, which it positions as its strongest model yet for writing code and for "cowork" — working alongside a person on multi-step tasks rather than answering one question at a time.

-
- Anthropic Releases Claude Opus 5 -
-
What this means
-

The pattern keeps repeating: what cost a premium six months ago costs half as much now. If you have held off on AI tools because the good ones were too expensive for real work — reviewing a long contract, working through a research task over hours rather than minutes — that calculation keeps moving in your favour.

+

The list of models that can credibly handle real work keeps growing, and Chinese labs are now a standing part of it rather than a curiosity. If your organisation is choosing an AI tool, the practical effect is more competition on price and fewer reasons to lock into one vendor.

Source - Anthropic ↗ + Qwen ↗
@@ -468,23 +464,23 @@

Anthropic Releases Claude Opus 5

@@ -507,16 +503,16 @@

What else happened today

▸ - ByteDance Releases Seedance 2.5 - ByteDance Seed + Californians Can Now Force Data Brokers to Delete Their Data + NBC 7 San Diego
-

ByteDance published Seedance 2.5, an update to its AI video model built around generating a shot in one take and using reference images to keep characters and settings consistent. Consistency has been the weak point of AI video — faces and props drift between clips — so that is the part of the problem worth watching.

+

California's DROP system — one request that reaches the data brokers holding your personal information — became enforceable on August 1. Californians can file a single deletion request instead of chasing hundreds of companies individually. If you work in marketing, sales, or anywhere that buys contact lists, expect the lists to start shrinking.

- VIA ByteDance Seed ↗ + VIA NBC 7 San Diego ↗

@@ -524,16 +520,20 @@

What else happened today

▸ - China Starts Making Its Own Advanced Chipmaking Machines - Reuters + A Working Example of GitHub's Stacked Pull Requests + GitHub Blog
-

China has begun producing home-grown immersion deep-ultraviolet lithography machines — the equipment that prints circuits onto silicon — according to a source cited by Reuters. Only a handful of companies outside China can build these, and export controls have kept them out of Chinese factories.

+

Previously covered on 2026-07-31: GitHub put stacked pull requests — splitting one large change into a chain of smaller reviewable pieces — into public preview. GitHub has now published a walkthrough of someone using the feature end to end to modernise an old codebase. It is the clearest picture yet of what the workflow actually looks like day to day.

+
+ A Working Example of GitHub's Stacked Pull Requests +
+

- VIA Reuters ↗ + VIA GitHub Blog ↗

@@ -541,16 +541,16 @@

What else happened today

▸ - AI Financial Advice Holds Up Better Than Expected - MIT Sloan + Why People Get Attached to Their Tools + Stack Overflow Blog
-

MIT Sloan looked at how well AI assistants answer personal finance questions and found the advice largely holds up — provided you ask specific questions rather than vague ones. The difference between a useful answer and a generic one comes down to how much of your actual situation you put in the question.

+

Stack Overflow published an argument that developers cling to particular tools because a tool people have used for years encodes trust — you know exactly how it behaves when things go wrong. It is a useful frame for anyone being asked to swap a familiar tool for an AI-assisted one: resistance is usually about predictability, not stubbornness.

- VIA MIT Sloan ↗ + VIA Stack Overflow Blog ↗

@@ -558,16 +558,16 @@

What else happened today

▸ - US Warns of Attacks on Water Utility Controllers - Censys + Thinking Machines Publishes a Model Called Inkling + Hugging Face
-

CISA, the US cybersecurity agency, issued an alert about attackers targeting the programmable controllers that run water treatment and distribution equipment. These are small industrial computers that were never designed to sit on the public internet, and plenty of them still do.

+

The AI lab Thinking Machines released a model named Inkling on Hugging Face, the site where open models are shared and downloaded. Details are thin so far, but it is another lab choosing to publish rather than keep its work behind an API.

- VIA Censys ↗ + VIA Hugging Face ↗

@@ -588,48 +588,56 @@

For the technically curious

Deep Dive
-

TLS 1.2's Oldest Key Exchange Methods Are Officially Deprecated

-

The IETF published RFC 10015, which formally deprecates obsolete key exchange methods in TLS 1.2 and DTLS 1.2 — the handshake step where a browser and a server agree on a shared secret before any encryption starts. Deprecation is not removal: the methods still function, but the standards body is now telling implementers to stop offering them.

+

Dependabot Now Waits Three Days Before Suggesting an Update

+

Dependabot is the GitHub service that automatically opens a pull request when a piece of third-party code your project depends on publishes a new version. GitHub changed the default so those version-update pull requests now wait three days after a release, giving maintainers and security researchers a window to spot problems in a release before it lands in your code. Security fixes are exempt and still go out immediately.

- - + +
Protocol versionStatus of these key exchange methods
TLS 1.2 / DTLS 1.2Deprecated by RFC 10015 — still functional, no longer recommended
TLS 1.3 / DTLS 1.3Never included — left out at design time
Update typeOld defaultNew default
Version updatePull request opened as soon as the release landsPull request waits 3 days
Security updateOpened immediatelyStill opened immediately
Why it matters
-

Most of what is on that list lacks forward secrecy, which means a server key stolen years from now could decrypt traffic captured today. TLS 1.3 left these methods out at design time; this closes the same door on the older version that much of the long tail of the internet still runs.

+

The old default made every project an early adopter by accident, including the bad releases. Three days is short enough that you stay current and long enough that most withdrawn or compromised releases get caught first — a small default change that quietly shifts risk off thousands of repositories.

Source - RFC Editor ↗ + GitHub Blog ↗
Deep Dive
-

A Soundness Bug in Lean's Proof Kernel, Written Up in Full

-

Leonardo de Moura published a postmortem for bug #14576 in the kernel of Lean, the proof assistant — the small trusted core that checks every proof the system accepts. A soundness bug there means the checker can be talked into accepting something false, which is the one category of bug a proof assistant cannot live with.

+

A Language Model Running on a 6502

+

A developer got an autoregressive language model — the same basic design as the chatbots you use, generating one word at a time and feeding each one back in — running on the 6502, the 8-bit processor from the 1970s that powered the Apple II, the Commodore 64 and the NES. It is a demonstration rather than a product: the machine has a few kilobytes of memory and runs at a fraction of a modern phone's speed.

Why it matters
-

Lean is increasingly used to verify mathematics and, in a few places, real software. That whole arrangement rests on the kernel being small enough to audit and trust, so the useful part of a postmortem like this is less the bug itself than what it shows about how that core gets scrutinised.

+

The interesting part is what it proves about the design rather than the hardware. The loop below is the whole idea behind text generation, and it is small enough to fit on a chip from fifty years ago — the intelligence comes from the size of the model's weights, not from any complexity in the surrounding machinery.

+ 
# The entire generation loop, in five lines
+tokens = [start_token]
+for _ in range(50):
+    scores = model(tokens)        # score every possible next token
+    next_token = sample(scores)   # pick one
+    tokens.append(next_token)     # feed it back in
+print(detokenize(tokens))
+
Source - Leonardo de Moura ↗ + Matt Beton ↗
@@ -643,7 +651,7 @@

A Soundness Bug in Lean's Proof Kernel, Written Up in
◇ Fun fact
-

"A 15-year-old designed and built a cycloidal gearbox, then posted the whole thing on GitHub."

+

"A working language model now runs on the 8-bit chip from the Commodore 64."

diff --git a/seen_topics.json b/seen_topics.json index ee12027..b3bce20 100644 --- a/seen_topics.json +++ b/seen_topics.json @@ -1,60 +1,4 @@ [ - { - "date": "2026-07-26", - "title": "Cloudflare Lets Websites Block or Charge AI Crawlers", - "summary": "Cloudflare, which sits in front of a large share of the world's websites, rolled out new controls that let site owners decide how AI companies' bots may use their content — allow them, block them, or charge them per visit.", - "source_urls": [ - "https://blog.cloudflare.com/content-independence-day-ai-options/" - ] - }, - { - "date": "2026-07-26", - "title": "Thinking Machines Releases Its First Model, Inkling", - "summary": "Thinking Machines, the high-profile startup founded by former OpenAI leaders, published its first model, Inkling, on Hugging Face.", - "source_urls": [ - "https://huggingface.co/blog/thinkingmachines-inkling" - ] - }, - { - "date": "2026-07-26", - "title": "Bitchat Resurfaces After India's Takedown Order", - "summary": "Previously covered on 2026-07-25: India ordered GitHub to remove Bitchat, Jack Dorsey's Bluetooth-based messaging app.", - "source_urls": [ - "https://radicle.network/nodes/rosa.radicle.network/rad%3Az2v9tRJz1oknFAqCSY5W5c76nVvm6" - ] - }, - { - "date": "2026-07-26", - "title": "A Top Law School Rethinks How It Teaches for the AI Era", - "summary": "The University of Chicago Law School published a strategy for how legal education should change now that AI can draft, research, and review documents.", - "source_urls": [ - "https://www.law.uchicago.edu/news/ai-strategy-statement" - ] - }, - { - "date": "2026-07-26", - "title": "GM Bets on Sodium Batteries for the Power Grid", - "summary": "General Motors is backing sodium-ion batteries — which swap out the scarce lithium found in most batteries for cheap, abundant salt — to store energy for the US electric grid.", - "source_urls": [ - "https://spectrum.ieee.org/sodium-ion-battery-peak-energy" - ] - }, - { - "date": "2026-07-26", - "title": "Running a 28.9M-Parameter Language Model on an $8 Chip", - "summary": "A developer got a small language model — 28.", - "source_urls": [ - "https://github.com/slvDev/esp32-ai" - ] - }, - { - "date": "2026-07-26", - "title": "Why Automatically Picking the Right AI Model Is Harder Than It Looks", - "summary": "IBM researchers dug into 'model routing' — the idea of automatically sending each question to the cheapest AI model that can still answer it well, and only escalating hard ones to an expensive model.", - "source_urls": [ - "https://huggingface.co/blog/ibm-research/model-routing-is-simple-until-it-isnt" - ] - }, { "date": "2026-07-27", "title": "A Phone Wiped Itself at the Airport. Now Its Owner Is Facing Charges.", @@ -446,5 +390,61 @@ "source_urls": [ "https://leodemoura.github.io/blog/2026-8-1-postmortem-for-kernel-soundness-bug-14576/" ] + }, + { + "date": "2026-08-03", + "title": "Alibaba's Qwen Team Releases Qwen3.8-Max", + "summary": "Alibaba's Qwen team released Qwen3.", + "source_urls": [ + "https://qwen.ai/blog?id=qwen3.8" + ] + }, + { + "date": "2026-08-03", + "title": "Californians Can Now Force Data Brokers to Delete Their Data", + "summary": "California's DROP system — one request that reaches the data brokers holding your personal information — became enforceable on August 1.", + "source_urls": [ + "https://www.nbcsandiego.com/nbc-7-responds-2/californians-data-deletion-requests-drop-become-enforceable-aug-1/4054771/" + ] + }, + { + "date": "2026-08-03", + "title": "A Working Example of GitHub's Stacked Pull Requests", + "summary": "Previously covered on 2026-07-31: GitHub put stacked pull requests — splitting one large change into a chain of smaller reviewable pieces — into public preview.", + "source_urls": [ + "https://github.blog/ai-and-ml/github-copilot/stacked-sessions-and-pull-requests-in-the-github-copilot-app/" + ] + }, + { + "date": "2026-08-03", + "title": "Why People Get Attached to Their Tools", + "summary": "Stack Overflow published an argument that developers cling to particular tools because a tool people have used for years encodes trust — you know exactly how it behaves when things go wrong.", + "source_urls": [ + "https://stackoverflow.blog/2026/07/29/developers-are-attached-to-tools-because-tools-encode-trust/" + ] + }, + { + "date": "2026-08-03", + "title": "Thinking Machines Publishes a Model Called Inkling", + "summary": "The AI lab Thinking Machines released a model named Inkling on Hugging Face, the site where open models are shared and downloaded.", + "source_urls": [ + "https://huggingface.co/blog/thinkingmachines-inkling" + ] + }, + { + "date": "2026-08-03", + "title": "Dependabot Now Waits Three Days Before Suggesting an Update", + "summary": "Dependabot is the GitHub service that automatically opens a pull request when a piece of third-party code your project depends on publishes a new version.", + "source_urls": [ + "https://github.blog/security/supply-chain-security/the-case-for-a-cooldown-why-dependabot-now-waits-before-issuing-version-updates/" + ] + }, + { + "date": "2026-08-03", + "title": "A Language Model Running on a 6502", + "summary": "A developer got an autoregressive language model — the same basic design as the chatbots you use, generating one word at a time and feeding each one back in — running on the 6502, the 8-bit processor from the 1970s that powered the Apple II, the Commodore 64 and the NES.", + "source_urls": [ + "https://mattbeton.com/blog/bitnet-6502.html" + ] } ] \ No newline at end of file