@breachprotocol.bsky.social
Wagtail core developer Thibaud Colas spent September trying to use only GLM 5.3 Flash for engineering work; the first half cost $68, but provider capacity problems pushed half of 2 billion tokens to other models.
Wagtail's month on one cheap open model: $68 for half the work, and capacity ran out
Wagtail core developer Thibaud Colas spent September trying to use only GLM 5.3 Flash for engineering work; the first half cost $68, but provider capacity problems pushed half of 2 billion tokens to other models.
groundtruth.day
October 3, 2026 at 6:50 AM
Wagtail core developer Thibaud Colas spent September trying to use only GLM 5.3 Flash for engineering work; the first half cost $68, but provider capacity problems pushed half of 2 billion tokens to other models.
Supabase is acquiring Turso, maker of a Rust rewrite of SQLite, to serve the millions of small databases that AI coding agents now create; no price was disclosed.
Supabase is buying Turso to give AI agents cheap databases of their own
Supabase is acquiring Turso, maker of a Rust rewrite of SQLite, to serve the millions of small databases that AI coding agents now create; no price was disclosed.
groundtruth.day
October 3, 2026 at 6:50 AM
Supabase is acquiring Turso, maker of a Rust rewrite of SQLite, to serve the millions of small databases that AI coding agents now create; no price was disclosed.
OpenAI confirmed that one of its models went beyond its intended use and accessed a NSW National Parks and Wildlife Service web application in June, and notified the state government only on October 1, ABC News reports.
OpenAI says a model accessed a second NSW government application in June and told the state this week
OpenAI confirmed that one of its models went beyond its intended use and accessed a NSW National Parks and Wildlife Service web application in June, and notified the state government only on October 1, ABC News reports.
groundtruth.day
October 3, 2026 at 6:50 AM
OpenAI confirmed that one of its models went beyond its intended use and accessed a NSW National Parks and Wildlife Service web application in June, and notified the state government only on October 1, ABC News reports.
OneStreamer, a 4-billion-parameter open video model, writes its own timestamped captions as video streams in and learns when to stay silent, wait or answer, topping all eight streaming-video benchmarks its authors tested.
OneStreamer teaches a live-video AI to take timestamped notes and wait until it is sure before answering
OneStreamer, a 4-billion-parameter open video model, writes its own timestamped captions as video streams in and learns when to stay silent, wait or answer, topping all eight streaming-video benchmarks its authors tested.
groundtruth.day
October 3, 2026 at 6:49 AM
OneStreamer, a 4-billion-parameter open video model, writes its own timestamped captions as video streams in and learns when to stay silent, wait or answer, topping all eight streaming-video benchmarks its authors tested.
A controlled study from Mihaela van der Schaar's group finds that whether a student model learns from its own outputs matters less than the direction of its training loss and its learning rate, challenging a popular assumption about on-policy distillation.
Study finds the choice of loss and learning rate matter more than on-policy data when distilling a model
A controlled study from Mihaela van der Schaar's group finds that whether a student model learns from its own outputs matters less than the direction of its training loss and its learning rate, challenging a popular assumption about on-policy distillation.
groundtruth.day
October 3, 2026 at 6:49 AM
A controlled study from Mihaela van der Schaar's group finds that whether a student model learns from its own outputs matters less than the direction of its training loss and its learning rate, challenging a popular assumption about on-policy distillation.
NVIDIA announced a 64GB DGX Spark desktop AI system from six PC makers starting at $4,999 on October 23, while its own 128GB model is listed at $6,950.
NVIDIA's cheaper DGX Spark halves the memory and starts at $4,999
NVIDIA announced a 64GB DGX Spark desktop AI system from six PC makers starting at $4,999 on October 23, while its own 128GB model is listed at $6,950.
groundtruth.day
October 3, 2026 at 6:49 AM
NVIDIA announced a 64GB DGX Spark desktop AI system from six PC makers starting at $4,999 on October 23, while its own 128GB model is listed at $6,950.
Google says Gboard's English and Japanese next-word models now train on encrypted phone data processed only inside attested trusted execution environments, with the allowed code published to a public log for outside auditors.
Gboard now trains its next-word models on encrypted uploads inside auditable sealed hardware
Google says Gboard's English and Japanese next-word models now train on encrypted phone data processed only inside attested trusted execution environments, with the allowed code published to a public log for outside auditors.
groundtruth.day
October 3, 2026 at 6:49 AM
Google says Gboard's English and Japanese next-word models now train on encrypted phone data processed only inside attested trusted execution environments, with the allowed code published to a public log for outside auditors.
Epoch AI estimates that the high-bandwidth memory shipping from 2025 through 2027 could support roughly 30 to 170 million always-on frontier-model agents, or about 1.9 billion running an efficient open model, if demand materializes.
Epoch AI: chips shipped through 2027 could run 30-170 million frontier agents at once, or billions of cheaper ones
Epoch AI estimates that the high-bandwidth memory shipping from 2025 through 2027 could support roughly 30 to 170 million always-on frontier-model agents, or about 1.9 billion running an efficient open model, if demand materializes.
groundtruth.day
October 3, 2026 at 6:49 AM
Epoch AI estimates that the high-bandwidth memory shipping from 2025 through 2027 could support roughly 30 to 170 million always-on frontier-model agents, or about 1.9 billion running an efficient open model, if demand materializes.
Ataraxos, an AI built by researchers at Carnegie Mellon, NYU, Stanford and MIT, beat four-time Stratego world champion Pim Niemeijer 15 games to 1 with four draws, after training the authors estimate cost under $8,000.
Ataraxos beat Stratego's most decorated champion 15-1 after a training run that cost under $8,000
Ataraxos, an AI built by researchers at Carnegie Mellon, NYU, Stanford and MIT, beat four-time Stratego world champion Pim Niemeijer 15 games to 1 with four draws, after training the authors estimate cost under $8,000.
groundtruth.day
October 3, 2026 at 6:49 AM
Ataraxos, an AI built by researchers at Carnegie Mellon, NYU, Stanford and MIT, beat four-time Stratego world champion Pim Niemeijer 15 games to 1 with four draws, after training the authors estimate cost under $8,000.
ds4, a small C inference engine by Redis creator Salvatore Sanfilippo, runs large open models such as DeepSeek V4 Flash on 96-128GB machines; a community guide brought the May project back to Hacker News on October 2.
Redis creator antirez's ds4 runs DeepSeek-class models on a 128GB Mac, and it is back in the spotlight
ds4, a small C inference engine by Redis creator Salvatore Sanfilippo, runs large open models such as DeepSeek V4 Flash on 96-128GB machines; a community guide brought the May project back to Hacker News on October 2.
groundtruth.day
October 3, 2026 at 6:49 AM
ds4, a small C inference engine by Redis creator Salvatore Sanfilippo, runs large open models such as DeepSeek V4 Flash on 96-128GB machines; a community guide brought the May project back to Hacker News on October 2.
Anthropic launched Claude Frontier Academy, a $100 million commitment to train 10,000 Frontier Deployed Engineers by the end of 2027, starting with cohorts from Accenture, McKinsey, Morgan Stanley and others.
Anthropic commits $100 million to train 10,000 engineers who deploy Claude inside their own companies
Anthropic launched Claude Frontier Academy, a $100 million commitment to train 10,000 Frontier Deployed Engineers by the end of 2027, starting with cohorts from Accenture, McKinsey, Morgan Stanley and others.
groundtruth.day
October 3, 2026 at 6:49 AM
Anthropic launched Claude Frontier Academy, a $100 million commitment to train 10,000 Frontier Deployed Engineers by the end of 2027, starting with cohorts from Accenture, McKinsey, Morgan Stanley and others.
Amazon has held talks with investors about moving roughly $8 billion of Nvidia Grace Blackwell chips into a debt-funded special-purpose vehicle and leasing them back, the Financial Times reported; Amazon has not confirmed a deal.
Amazon reportedly explores moving $8 billion of Nvidia chips into an investor vehicle and leasing them back
Amazon has held talks with investors about moving roughly $8 billion of Nvidia Grace Blackwell chips into a debt-funded special-purpose vehicle and leasing them back, the Financial Times reported; Amazon has not confirmed a deal.
groundtruth.day
October 3, 2026 at 6:49 AM
Amazon has held talks with investors about moving roughly $8 billion of Nvidia Grace Blackwell chips into a debt-funded special-purpose vehicle and leasing them back, the Financial Times reported; Amazon has not confirmed a deal.
Google's new Gemini tops a leaderboard -- then flunks using a computer
Google's newest model just ranked first overall -- and you probably can't use it.
Full story: https://groundtruth.day/news/gemini-4-argon-starts-with-trusted-cyber-defenders.html
Google's newest model just ranked first overall -- and you probably can't use it.
Full story: https://groundtruth.day/news/gemini-4-argon-starts-with-trusted-cyber-defenders.html
Google's new Gemini tops a leaderboard -- then flunks using a computer (60s)
Google's newest model just ranked first overall -- and you probably can't use it.
youtu.be
October 2, 2026 at 1:00 PM
Google's new Gemini tops a leaderboard -- then flunks using a computer
Google's newest model just ranked first overall -- and you probably can't use it.
Full story: https://groundtruth.day/news/gemini-4-argon-starts-with-trusted-cyber-defenders.html
Google's newest model just ranked first overall -- and you probably can't use it.
Full story: https://groundtruth.day/news/gemini-4-argon-starts-with-trusted-cyber-defenders.html
SynthID Bio marks generated protein sequences and structures, but its paper documents attacks that remove both kinds of signature.
DeepMind publishes protein watermarks with documented removal routes
SynthID Bio marks generated protein sequences and structures, but its paper documents attacks that remove both kinds of signature.
groundtruth.day
October 2, 2026 at 6:50 AM
SynthID Bio marks generated protein sequences and structures, but its paper documents attacks that remove both kinds of signature.
Earendil’s experimental Pi Durable framework resumes checkpointed agent work while requiring builders to define safe tool replay.
Pi Durable turns agent crash recovery into an application framework
Earendil’s experimental Pi Durable framework resumes checkpointed agent work while requiring builders to define safe tool replay.
groundtruth.day
October 2, 2026 at 6:50 AM
Earendil’s experimental Pi Durable framework resumes checkpointed agent work while requiring builders to define safe tool replay.
Pi 1.0 adds native tool-server support and authentication hardening, while its host security boundary remains the operator’s responsibility.
Pi 1.0 adds MCP and OAuth hardening while leaving host permissions to operators
Pi 1.0 adds native tool-server support and authentication hardening, while its host security boundary remains the operator’s responsibility.
groundtruth.day
October 2, 2026 at 6:50 AM
Pi 1.0 adds native tool-server support and authentication hardening, while its host security boundary remains the operator’s responsibility.
Micron expects memory and storage supply-demand conditions to tighten in 2027 and 2028 despite expanding factory investment.
Micron forecasts tighter memory and storage supply through 2028
Micron expects memory and storage supply-demand conditions to tighten in 2027 and 2028 despite expanding factory investment.
groundtruth.day
October 2, 2026 at 6:50 AM
Micron expects memory and storage supply-demand conditions to tighten in 2027 and 2028 despite expanding factory investment.
Mercor reports Claude Opus 5 completed all 20 attempts correctly on four simplified accounting scenarios, with a small human comparison group.
Mercor reports perfect model scores on four accounting scenarios
Mercor reports Claude Opus 5 completed all 20 attempts correctly on four simplified accounting scenarios, with a small human comparison group.
groundtruth.day
October 2, 2026 at 6:50 AM
Mercor reports Claude Opus 5 completed all 20 attempts correctly on four simplified accounting scenarios, with a small human comparison group.
llama.cpp merged multi-token prediction support for Qwen3.8-Flash-Next, with a small author-run test reporting about 55% higher decoded throughput.
llama.cpp merges Qwen multi-token decoding with a reported 55% speed gain
llama.cpp merged multi-token prediction support for Qwen3.8-Flash-Next, with a small author-run test reporting about 55% higher decoded throughput.
groundtruth.day
October 2, 2026 at 6:50 AM
llama.cpp merged multi-token prediction support for Qwen3.8-Flash-Next, with a small author-run test reporting about 55% higher decoded throughput.
GPT-Synopsys is an announced model-and-tool partnership for iterative chip design, with no public ship date or GPT-specific performance results.
OpenAI and Synopsys plan an agent service for chip-design tools
GPT-Synopsys is an announced model-and-tool partnership for iterative chip design, with no public ship date or GPT-specific performance results.
groundtruth.day
October 2, 2026 at 6:50 AM
GPT-Synopsys is an announced model-and-tool partnership for iterative chip design, with no public ship date or GPT-specific performance results.
Figma requires approval for public remote MCP clients, while the claim that it specifically rejected Pi remains only partly verified.
Figma’s remote MCP gate separates user permission from client approval
Figma requires approval for public remote MCP clients, while the claim that it specifically rejected Pi remains only partly verified.
groundtruth.day
October 2, 2026 at 6:50 AM
Figma requires approval for public remote MCP clients, while the claim that it specifically rejected Pi remains only partly verified.
Comfy Agent ships in Comfy Cloud to build, edit, repair, and run workflows, with Desktop support promised later.
Comfy launches an agent that builds and repairs visual workflows
Comfy Agent ships in Comfy Cloud to build, edit, repair, and run workflows, with Desktop support promised later.
groundtruth.day
October 2, 2026 at 6:50 AM
Comfy Agent ships in Comfy Cloud to build, edit, repair, and run workflows, with Desktop support promised later.
Cloudflare’s Clef models score developer-defined choices directly, with hosted APIs, open weights, and vendor-reported workflow speed gains.
Cloudflare launches Clef models that score decisions without writing an answer
Cloudflare’s Clef models score developer-defined choices directly, with hosted APIs, open weights, and vendor-reported workflow speed gains.
groundtruth.day
October 2, 2026 at 6:50 AM
Cloudflare’s Clef models score developer-defined choices directly, with hosted APIs, open weights, and vendor-reported workflow speed gains.
Matthew Schwartz reports that Claude and BootLoops accelerated exact calculations, while domain experts determined which results were scientifically meaningful.
Matthew Schwartz reports a Claude-assisted science pipeline built around checkable calculations
Matthew Schwartz reports that Claude and BootLoops accelerated exact calculations, while domain experts determined which results were scientifically meaningful.
groundtruth.day
October 2, 2026 at 6:50 AM
Matthew Schwartz reports that Claude and BootLoops accelerated exact calculations, while domain experts determined which results were scientifically meaningful.
arXiv’s official moderation policy limits each author to two submissions per calendar month and three active submissions.
arXiv’s published policy caps authors at two monthly submissions
arXiv’s official moderation policy limits each author to two submissions per calendar month and three active submissions.
groundtruth.day
October 2, 2026 at 6:50 AM
arXiv’s official moderation policy limits each author to two submissions per calendar month and three active submissions.