Atria Dawn 744B Launch, Qualcomm-AWS AI Chip Deal, and More
MIT-licensed 744B agentic model, a $60B Qualcomm-AWS chip partnership, GPU-alternative startup Euclyd raising $230M, and other key tech stories from the third week of September 2026
BRAIN OS INSTITUTE Engineering Blog
Notes from building BOSI and IPJP — brain-cognition assessments that classify 8,192 types. Engineering, research, and product.
Latest
Earlier posts
AI Pacing Debate Heats Up, iOS 27 Launches, and More Tech Headlines
Dario Amodei calls for slowing AI development, Microsoft publishes MAI rules, SoftBank secures $11.87B for OpenAI, South Korea chips shatter export records, and Marvel Wolverine drops on PS5.
GPT-Live API, Asimov Inference Chip, and Key Tech News This Week
OpenAI voice AI API, Salesforce job-ready agents, Positron AI challenging NVIDIA inference, and more highlights from this week in tech.
AI Agent Security Alerts, TSMC Revenue Record, and Key Tech News
Sandbox escape vulnerabilities found in 7 AI coding agents, AI-powered mass PaperCut hack, Meta acquires Stilla.ai, and TSMC posts all-time high monthly revenue.
Two filled on try four — Qwen3.8-27B 4bit on Terminal-Bench 89, pass@4 48 and the vanished ERRs
Round 4 of Qwen3.8-27B MLX 4bit (17GB): restart only the server, rerun the 43 failures. Two recoveries reach pass@4 48/89 and the 13 round-3 ERRs drop to 1 — the server was never the cause.
NVIDIA Acquires Hugging Face, OpenAI Agents API Launch, and More
NVIDIA closes a $12.9B Hugging Face deal, OpenAI ships the Agents API beta, Oracle cloud grows 121% — the tech stories worth your attention this week.
Third try recovered nothing — Qwen3.8-27B 4bit on Terminal-Bench 89, pass@3 46 and the 13 ERRs
Round 3 of Qwen3.8-27B MLX 4bit (17GB): re-running the 43 round-2 failures recovers zero, pass@3 stays 46/89. 30 failed exactly as before; 13 never ran a turn — and the real cause was not the server.
Google's 13B Finland Nuclear Deal, Harvey AI $550M Round, and More
Google locks in a 22-year nuclear power deal for Finnish AI data centers, Harvey AI nearly doubles its valuation, NSA warns of Chinese model distillation, and Nintendo Direct drops major reveals.
How much does a retry recover — Qwen3.8-27B 4bit on Terminal-Bench 89, pass@2 46 vs 53
Round 2 of Qwen3.8-27B MLX 4bit (17GB): re-running the 54 round-1 failures reaches pass@2 46/89 (51.7%). 11 recoveries (median 6.5 min), two solved only by 27B, and why retrying failures is slower.
Meta Muse Launch, OpenAI Solves Millennium Prize Problem, and More
Meta launches Muse personal AI agent, OpenAI claims Navier-Stokes solution, Cognition AI hits $48B valuation, Qualcomm-AWS seal AI chip deal, and other key tech stories for September 10.
The price of a 6x smaller model — Qwen3.8-27B 4bit on Terminal-Bench 89, single pass 35 vs 45
Qwen3.8-27B MLX 4bit (17GB) on Terminal-Bench 2.1, same harness as Flash Next (100GB): 35/89 (39.3%) vs stock 45 and Uncensored 41. Which tasks split, 15 repetition loops, speed, ops notes
Apple Foldable iPhone Ultra, Mistral AI Raises Record 3B Euros, ASML 12-Inch Mask Shift
Apple is set to unveil its first foldable iPhone Ultra, Mistral AI closes the largest European tech funding round at 3B euros, and ASML unites chip giants behind 12-inch photomasks.
Closing the table — stock vs Uncensored Qwen3.8 on Terminal-Bench 89, pass@5 64 to 59
Round-5 (pass@5) results for the abliterated Uncensored Qwen3.8 Flash Next on Terminal-Bench 2.1, plus a series wrap-up: 59/89 vs stock 64/89, the 13 tasks that split them, speed, and harness failures
SoundHound-LivePerson Merger, Foxconn AI Revenue Surge & More
SoundHound AI closes the LivePerson deal, Foxconn AI servers hit 50% of revenue, CISA flags LiteLLM MCP bypass, Pixxel raises $100M, and Onimusha sells 1M on launch day.
NVIDIA Buys Hugging Face, Claude Proves Fermat Theorem
NVIDIA acquires Hugging Face for $12.93B, Claude formalizes Fermat Last Theorem in Lean, and Apple gets a new CEO. Key tech stories this week.
Round five: the "saturated" curve bent up again — stock Qwen3.8 pass@5 64/89 (71.9%)
Round-5 (pass@5) Terminal-Bench 2.1 results for stock Qwen3.8 Flash Next mixed-4-8bit: 4 of 29 retries recovered (two timeout-with-the-right-file), a 120K context handoff and a fallback loop
Giving the other side the same number of tries — Uncensored Qwen3.8 pass@4 58/89 (65.2%), gap of 6
Round-4 (pass@4) Terminal-Bench 2.1 results for the abliterated Uncensored Qwen3.8 Flash Next: 2 of 33 retries recovered, two fallback loops worth 3,877 HTTP 400s, and a 40-minute Coq build
AI Agent Security Funding Surge, NVIDIA NVHBM, and Gaming Big Week
From HiddenLayer $100M and Gimlet Labs $300M AI raises to NVIDIA custom HBM, PlayStation State of Play, and Nintendo Direct — this week in tech.
Still ahead after three tries: stock Qwen3.8 pass@3 58/89 vs Uncensored best-of-3 56
Round-3 (pass@3) Terminal-Bench 2.1 results for stock Qwen3.8 Flash Next mixed-4-8bit, a like-for-like three-trial comparison with the abliterated build, and why 31 tasks still fail
Four tries, still ahead: stock Qwen3.8 pass@4 60/89 (67.4%) and the two tasks that flipped
Round-4 (pass@4) Terminal-Bench 2.1 results for stock Qwen3.8 Flash Next mixed-4-8bit: 2 recoveries out of 31 retries, and why 29 still fail — repetition loops, premature submits, near misses
Crusoe Hits $30B, Tesla Cybercab Under NHTSA Probe, OpenShot 4.0
Crusoe triples valuation to $30B, Tesla Cybercab draws an NHTSA probe on launch day, Microsoft ships MAI-Transcribe-2 at $0.10/hr, and more from this week in tech.
Fable 5.1 Launch, Meta $18B Settlement, AWS DuckLabs Deal
Anthropic ships Fable 5.1 with 75% cheaper cache reads, Meta pays up to $18B in child safety settlement, and AWS acquires DuckLabs while DuckDB stays MIT-licensed.
Does abliteration boost coding? Uncensored vs stock Qwen3.8 on Terminal-Bench 89
A same-conditions single-pass A/B between an abliterated (Uncensored) model and the stock mixed-4-8bit build across all 89 Terminal-Bench 2.1 tasks
OpenAI Astra Hits Critical Risk, NVIDIA DLSS 5 Launches, Uber Cuts 3,300 Jobs
OpenAI flags Astra as its first critical-risk cyber model, NVIDIA ships DLSS 5, and Dell books $60.9B in AI orders. Key tech stories from early September 2026.
Uncensored Qwen3.8 Flash Next: all 89 Terminal-Bench tasks
Three days of Terminal-Bench 2.1 on a community abliterated derivative. Four rounds of context, timeout and temperature changes, 13 server wedges, and what a final 56/89 actually means.
EU classifies ChatGPT as a search engine, Runway unveils Solaris, FTC sues Amazon
The EU designated ChatGPT a very large online search engine and the FTC sued Amazon over hidden ad surcharges. From Runways Solaris to Europes new AI supercomputer, the weeks notable stories.
Running Terminal-Bench overnight against Qwen3.8 Flash Next
Measuring how much real terminal work a local LLM on an M5 Max can actually complete. The pass rate was less interesting than the server wedging four times and recovering itself every time.
Apple gets a new CEO, DeepSeek Harness hits 112K stars, and other tech news
John Ternus takes over as Apple CEO ahead of the September event, DeepSeek releases its Harness agent framework, six CVEs land in ash_ai, plus OpenMAIC and the ArcBox isolation runtime.
Qwen3.8 27B — 4bit vs 8bit vs Flash Next, and is DFlash2 really lossless?
Measuring Qwen3.8 27B at 4bit and 8bit against Flash Next, then adding DFlash2 speculative decoding. It was 2–3x faster, but contrary to the official claim the output diverged from the original on M5.
Running Qwen3.8 Flash Next on an M5 Max, and the fight with prefill
Running a 125B MoE model on one Mac as an agent backend. I assumed the model was slow. A day of digging showed the real culprit was prefill — and how caching and a few flags fixed it.
Apple ships M6, WRTN becomes a unicorn, Gatik raises $200M, and more
Apples 2nm M6 chip, Intels 256-core Diamond Rapids, WRTN becoming Koreas first AI service unicorn, and Emerald AIs data center power optimization — nine stories from the last week of August.
AI productivity skepticism, German funding for Flatpak, and other tech news
An NBER survey finding 89 percent of executives see no productivity effect from AI, Germanys investment in Flatpak, and the KDE Gear 26.08 release — this weeks notable stories.
Anthropic wins against the Pentagon, Instinct AI surges, and other tech news
A court rules Anthropics blacklisting unconstitutional, the A2A standard consolidates, Instinct AI raises $250M, and sched_ext lands complete in Linux 7.3 — eight stories from the week.
Buzz — the open-source workspace that gives agents an identity
Block released Buzz, which treats AI agents as workspace members rather than add-ons. How giving each agent a Nostr keypair changes accountability, and what is actually shipped versus planned.
NVIDIA posts $96.2B, OpenAI unveils its own chip, and other tech news
NVIDIA reported record quarterly revenue and moved on Hugging Face while OpenAI claimed its in-house inference chip beats Blackwell. A roundup of the weeks major moves across the industry.
[postgresql] pg_wal that will not shrink — start with replication slots
PostgreSQL will not delete WAL while a replication slot still holds it. Two months of WAL accumulated after a standby died, and the secondary damage that only surfaced once the space came back.
Career Counseling — BOSI for adults and teams
What changes when the same brain-cognition assessment is applied to working adults and companies instead of students. Career design, hiring, team composition, and finding a co-founder.
Edu Counseling — BOSI, rebuilt for parents
Our BOSI brain-cognition assessment, packaged so that parents of school-age children can act on it. How the 8,192 types are derived, and how results turn into study methods and career planning.
Starting an engineering blog
We build brain-cognition assessments that sort people into 8,192 types. This is where we write down the problems we hit along the way and the choices we made.
[proxmox] Fixing "CPU does not support x86-64-v2"
Modern container images require x86-64-v2. The default CPU model your hypervisor hands a VM does not have those instructions, so the process dies inside glibc before the application ever runs.
Fixing ssh "Connection timed out during banner exchange"
If your iptables whitelist only inspects ctstate NEW, the handshake never completes. A port scan still reports the port open, which is why the cause of a failing backup went unfound for a month.
[haproxy] The reload succeeded and the config still did not change
A reload can return rc=0 without applying anything. Two patterns: an old process surviving and serving the previous config, and changes landing in a staging file that never got promoted.
Recovering a failed Ubuntu 22.04 → 26.04 upgrade without reinstalling
Skipping an LTS step and stalling at the configure stage takes sudo and networking with it. Counting dpkg states tells you whether recovery without a reinstall is possible.
[monitoring] When "last updated at" is the wrong thing to alert on
If you alert on a value that does not change while things are working, it will fire eventually — guaranteed. The metric that produced false alarms, and the real alert we silenced while fixing it.
[linux] You changed the timezone, but the DB and containers did not
timedatectl changes the OS timezone. Your database and your containers keep the old one. How NOW() came back nine hours apart inside one cluster, and why nothing short of a restart fixes it.