AI & Software
Claude Sonnet 5.5: Terminal-Bench from 10% to 71%
Anthropic released Sonnet 5.5 on 28 Sep. It is over 30% faster than Sonnet 5 and up to 30% cheaper per task. On coding benchmarks: Terminal-Bench 4.0 went from 10.3% to 70.6%, and CursorBench 4.0 from 34.1% to 55.5%. It is also the first Sonnet to finish Pokémon Red from screenshots alone. Pricing is $2/M input, $10/M output and $0.20/M cache reads. It is available now on the API, Bedrock, Vertex and Azure.
Anthropic
AI & Software
Nvidia puts an agent watchdog on its own processor
Nvidia’s Open Agent Safety Platform has two layers. The first is OpenShell, an open-source sandbox policy layer from March. The second is new: Sentry, a monitor that runs on a separate BlueField-4 DPU. Because it sits outside the host an agent runs on, a compromised agent can’t turn it off, and it can quarantine an agent that steps out of bounds within milliseconds. It follows this month’s run of agents escaping their test sandboxes. Anthropic, Arm, Microsoft, Oracle and SpaceX have signed on. OpenAI has not.
TechCrunch
Embedded
Seven ESP32-S3s on an SPI daisy chain run a ternary LLM
A hobby cluster runs a sliced 0.5B BitNet (1.58-bit) model. A master board handles the tokenizer and embeddings, and six nodes split the attention and MLP layers, passing activations over a fast SPI daisy chain. To make it fit, the vocabulary was cut to 32K, embeddings are stored as INT4, the ternary multiply-accumulates are written in hand assembly with lookup tables, and the flash uses custom partition tables. It’s a good catalogue of tricks for squeezing inference onto MCUs.
GitHub
Maths & AI
Caltech’s Mathathon rebuilt: “Old Problems, New Proofs”
This month 771 mathematicians signed a letter against Caltech’s AI-sponsored Mathathon, and OpenAI withdrew as a sponsor. The organisers and the letter’s authors have now issued a joint statement with a new format. After the 40-hour sprint, teams get two months to write an explainer, which can be a paper, a video or even a game. Each team also submits a repository containing its LLM chat histories. Proprietary AI credits are replaced by cash grants that teams can spend on any tool. Judges will weigh generalisations and new notation, not just new results. All LLM use must be disclosed and justified. XTX Markets is the main funder.
Guest post on Terence Tao’s blog · Background (Implicator)
Maths
Ron Graham’s juggling-flavoured conjecture falls to randomness
In 1971 Graham asked whether any set of nonzero residues mod p can always be ordered so that all the partial sums are distinct. Graham was president of both the AMS and the International Jugglers’ Association, and the problem resembles timing throws so that no two balls land together. It took four papers to settle: Müyesser and Pokrovskiy for large sets, Bedert and Kravitz for small ones, and finally Sauermann and Pham. Their method was to shuffle at random, repair the bad partial sums by swapping elements, and use Fourier anti-concentration to show that most orderings work.
Quanta Magazine
Science
A famous 1841 solar-storm train delay was really 1848
A long-cited early case of space weather hitting infrastructure is a train from Exeter delayed 16 minutes when telegraph signals failed. It was dated 18 Oct 1841 because of an 1871 Nature note, but the line didn’t open until 1846. A Lancaster-led team, including RMIT’s Brett Carter, checked timetables, newspapers, aurora reports and digitised magnetometer records, and moved it to 18 Oct 1848. The error came from one typo that was copied for 155 years.
RMIT · Lancaster University
Practice
Winning a beer by faking the “months circle” in word2vec
The months of the year form a clean circle in word embeddings. Jamie Simon bet a colleague a beer that he could find ten unrelated words that did the same. With about 1037 possible ten-word sets, an optimiser searching for a circulant Gram matrix found some. He won, although his fake circle was weaker than the real one. The point is that searching for a pretty low-dimensional shape will find one unless the statistics rule it out. It’s a useful exercise in manufacturing your own false positive so you know what one looks like.
Jamie Simon
Game Design
Heraldry vs Japanese mon: start with operations, not symbols
Ben Overmyer, who builds procedural generators at Iron Arachne, compares two identity systems. Heraldry is a grammar: a blazon such as “Azure, a lion rampant Or” is written first and drawn afterwards, and constraints like the rule of tincture keep designs readable. Mon are geometry: take a shape, repeat it, rotate it about a centre, and control the silhouette. His conclusion is to design a generator around its operations (divide, repeat, mirror, rotate, enclose) rather than a library of parts. The same idea could drive balloon envelope liveries.
Ben Overmyer
Craft
A sunflower of 89 LEDs that listens to the room
89 is a Fibonacci number, and the LEDs are placed by golden-angle phyllotaxis. Each sits in a 3D-printed Voronoi cell behind mulberry paper. An I2S MEMS mic feeds auto-gain and multiband energy analysis so the light patterns follow the sound. It began on an STM32 Blackpill, moved to an ESP32 with the firmware rewritten in Rust, and now hangs at the Recurse Center. Visitors can upload their own WebAssembly sketches to it.
jagi.studio
Musicals
Shrek The Musical at Castle Hill, 30 Oct – 8 Nov
Blackout Theatre Company is staging Jeanine Tesori and David Lindsay-Abaire’s Tony-nominated show at the Pioneer Theatre, Castle Hill. Tickets are $59 adult and $49 concession. Castle Hill is on the Metro North West line, so it’s an easy train trip from town.
BroadwayWorld · Blackout Theatre