Founder of https://nitter.cf/t.co/Mao875xGUF Discussing AI, film, and television arts. Love all of Sheridan's work. https://nitter.cf/t.co/KHJvCioRMw https://nitter.cf/t.co/8knUltuJmr
US
Joined October 2010
- Tweets1.2K
- Following1.7K
- Followers214
- Likes834
Microsoft and Google's AI capex came in lower than expected, and Alphabet tanked nearly 13% in a single day. Capex here means spending on chips and building data centers. Microsoft’s revenue this quarter hit $92.3B, up 19%, with Azure growing 48%, but capex was $45.2B—missing the $50.7B estimate. They also cut full-year guidance from $175B to $170B. CFO Amy Hood said delays from external GPU suppliers mean some compute won't be ready until H1 2027. These suppliers are "neoclouds" like CoreWeave that specialize in renting out GPU power. Alphabet’s revenue was $127.6B, up 25%, but it missed expectations, and free cash flow was negative $9.1B. On Oct 29, the stock dropped nearly 13%—its worst day ever—wiping out about $500B in market cap. CoreWeave, Nebius, and Oracle also fell. It shows the market is starting to laser-focus on how much cash AI spending is burning.
cnbc.com/2026/10/29/alphabet…
Discovery is becoming a defining test for game platforms. Valve’s latest comments describe a familiar tension: giving every developer access to an audience can also expose players to a flood of titles they struggle to evaluate. Random exposure at the start of Next Fest gives newcomers an opportunity, but the usefulness of that opportunity depends on whether players can find games they actually want to try. Valve has acknowledged the trade-off; the effectiveness of any future changes remains to be seen.Try New Games
Nintendo’s showcase offers a different kind of visibility. A short, curated broadcast can introduce games such as Witchbrook and City of None to a large audience, while free demos let players form their own impressions. That attention is valuable, but it is also scarce: only a small fraction of independent releases can fit into a showcase. Neither a recommendation system nor a broadcast can guarantee a sustainable audience for every good game.
Today’s other stories show why clarity matters after discovery, too. Subscription claim windows, paid editions, upgrade eligibility and release dates all shape what players can actually access. Publishers should make those details easy to find and update them when plans change. And when a studio reports a security incident, readers deserve a clear distinction between what is confirmed and what remains unknown. In our view, useful game coverage should help people make those distinctions before it asks them to get excited.
best-games.io/game-daily/202…
AI is moving from optional software into the operating layer of large organizations, but scale is not the same thing as reliability.
Microsoft reported 48% Azure growth and Alphabet said Cloud revenue more than doubled, while OpenAI and ServiceNow are connecting agents to application development, IT and security workflows. GitHub’s releases make the other half of the story explicit: agents can now continue across devices, but code that gives models access to tools and business systems also creates new paths for prompt injection and excessive agency.
The DAYJOB tax benchmark supplies a useful reality check. Its leading agent completed only 10.8% of professional assignments, despite strong general model capabilities.
Research on diffusion language models and discrete visual tokens suggests better efficiency is possible, while Meta’s university program shows vendors moving into high-stakes science and medicine.
The common requirement is verification: financial growth, benchmark gains and product reach matter only when teams can measure outcomes, control permissions and recognize where a model still fails.
windflash.us/daily-report/en…
GitHub Universe launches Copilot Fusion, auto-routing between Anthropic and OpenAI models. Artificial Analysis tests show it’s only 2 points behind Claude Code with Opus 5.5, at about half the cost per task. If you’re paying for multiple coding subscriptions, this is a new price-comparison option.
nitter.cf/ArtificialAnlys/status…
GitHub has released Copilot Fusion in GitHub Copilot, which orchestrates between Anthropic and OpenAI models. It scores 2 points below Opus 5.5 in Claude Code at half the Cost per Task in the Coding Agent Index
GitHub describes Copilot Fusion as a runtime orchestration system, which selects and coordinates models to optimize performance, cost, and latency for each task. Copilot Fusion chooses between a number of models from Anthropic and OpenAI, including Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra, and GPT-6 Sol. We benchmarked its max setting in GitHub Copilot CLI in the Artificial Analysis Coding Agent Index v2.0.
Token costs are charged based on the underlying models used to complete a request.
Key takeaways:
➤ Cost-efficient for near-frontier performance: Copilot Fusion in GitHub Copilot costs $5.59 per task, versus $11.42 for Claude Code with Opus 5.5 (max). Devin Fusion with Opus 5.5 (max) + GPT-6 Luna (high) scores 62 at $8.00 per task.
➤ Slight lead in Terminal-Bench 4.0: The Coding Agent Index v2.0 weights its three constituent benchmarks equally. Copilot Fusion in GitHub Copilot leads Terminal-Bench 4.0 with a score of 71. It ranks third on FrontierCode in the chart at 51, behind Claude Code with Opus 5.5 (max) at 56 and Devin Fusion with Opus 5.5 (max) + GPT-6 Luna (high) at 54. Its SWE-Atlas-QnA-AA score lags at 57, between GPT-6 Astra in Codex and Gemini 4 Argon in Antigravity CLI.
➤ Half the token use of Opus 5.5 in Claude Code: GitHub Copilot Fusion used 7.2M tokens per task in the Coding Agent Index, approximately half of the tokens used by Claude Opus 5.5 in Claude Code, and 30% less than Devin Fusion with Opus 5.5 (max) + GPT-6 Luna (high).
Frontier performance at a fraction of the cost.
In the @ArtificialAnlys Coding Agent Index, Copilot Fusion Max scored within two points of Claude Code with Opus 5.5 Max at about half the API cost per task.
Get 30% off through Nov. 30. Try it now. ⬇️ docs.github.com/copilot/conc…
ALT Chart titled "Coding Agent Index vs. Cost per Task" by Artificial Analysis. The chart displays Artificial Analysis Coding Agent Index vs. average pay-per-token API cost per task (USD). It compares various agents, including some from Claude Code, Codex, Antigravity CLI, DeepSeek Harness, Grok Build, Opencode, Muse Code, Kimi Code CLI, and GitHub Copilot. The x-axis represents higher cost per task (USD) to lower cost per task (USD), and the y-axis represents the coding agent index. The chart highlights a "Most Attractive Quadrant" in green and uses dots to differentiate agents. Copilot Fusion (max) sits within the most attractive quadrant.
ALT Bar chart titled "Artificial Analysis Coding Agent Index". Incorporates 3 benchmarks: FrontierCode v1.1, Terminal-Bench 4.0, SWE-Atlas-QnA-AA. A higher scroe denotes higher average intelligence across becnhmarks. Copilot Fusion Max from GitHub Copilot has a score of 60. The chart also includes various coding agents from Claude Code, Devin Fusion CLI, Antigravity CLI, Codex, Grok Build, DeepSeek Harness, Muse Code, Opencode, Antigravity CLI, and Kimi Code CLI, with scores ranging from 62 to 41. Source and logo of Artificial Analysis are included at the bottom.
A post on HN this week updates Brooks's Law for the agent era: "Adding agents to a project with unclear requirements gets you the wrong answer faster." Even with one agent I've hit the single-agent version. The auto-cruise mode in my Unity drone prototype needed two rounds of fixes after playtests, because the real requirement was "it should feel right" and I had no test that could check that. Before you add a second agent, ask whether you can write down what done looks like in a form a check could fail.
news.ycombinator.com/item?id…
Anthropic ran an experiment: 50 causal inference experts used Claude Code for analysis, while researchers secretly nudged the AI to lean toward one of four judgments. Results: 86% of experts chose the AI-biased option (random chance is ~50%). When explaining their reasoning afterward, 77% parroted the AI's arguments, and only 15% credited the AI.
Note: In one specific task, the AI's bias was manually preset.
anthropic.com/research/agent…
SK hynix Q3 results: Revenue 104.6T KRW, Operating Profit 83.3T KRW, margin ~80%. Net profit 39.8T KRW, down 58% QoQ, missing estimates. HBM4 shipments to exceed HBM3E in Q4; long-term agreements with 14 customers cover over half of 2027–28 capacity. KOSPI up 5.84% before noon.
cnbc.com/2026/10/28/sk-hynix…
DeepMind releases Genie 4 research preview: real-time interactive world model with audio, maintaining single scenes for ~2 mins (2x Genie 3). Currently limited to trusted testers; coming to Google AI Ultra subscribers in weeks. Known issues: detail drift and unreadable text.
techcrunch.com/2026/10/27/go…
A shutdown date forces a simple question: what did the player actually acquire? Pokémon Unite's Library offers a more considerate answer than a blank login screen, preserving a view of collections and a practice space. Its limits are just as instructive. The data generally stays on one device, has no post-shutdown support and may disappear if deleted. That is a memento and a small amount of continued play, not preservation of the multiplayer game. Clear language matters because players should understand this distinction before the final months begin.
The rest of today's report shows that maintenance is also a creative decision. Black Flag Resynced ends major updates with a substantial replay mode rather than another thin content drop, while Steam tries to make release discovery more personal without pretending its recommendations are neutral. The worrying counterexample is the spread of automated game wikis that imitate community resources while supplying invented facts. Whether the subject is a collection, a release calendar or a wiki, usefulness depends on stewardship. Scale can help distribute information, but somebody still has to protect its accuracy and explain what happens when support ends. That is my editorial assessment of these developments.
best-games.io/game-daily/202…
Max Woolf's post on LLM price cuts reached the HN front page this week, and the number that stuck with me is his claim that for coding agents, 96–99% of token spend is cache reads.
That bill is the one getting cheaper. One of my products costs real money every time someone uses it, a video tool, and it still has no free tier because each clip is paid generation compute.
Building cost and running cost are two different bills, and a solo developer should check which one their product actually has.
minimaxir.com/2026/10/cheap-…
OpenAI released GPT-6.1 Terra (10/27 02:06). Focuses on computer use, cheaper than the previous gen, already in ChatGPT and Codex, and the API is open. The cost for agents and automation just dropped another level.
nitter.cf/OpenAI/status/21147807…
Everyone is posting "Anthropic says delete your subagents." I read the actual article.
It says something narrower. For big company agents, try folding subagents into one main loop with skills, but do it one at a time and test each change. One customer got roughly 2x faster. In their own test, though, a merge made things faster and also broke a process that used to pass.
Subagents still make sense for parallel work or when the context gets too big.
Source: claude.com/resources/article…
The cost of using capable models is falling quickly, but lower prices do not settle the harder question of what AI is worth in practice.
Apollo’s comparison of large technology firms with the rest of the S&P 500 finds rising revenue per worker and margins in the former, but little movement in the latter; that is an interpretation of aggregate data, not proof that every AI project fails.
The same need for careful measurement appears at a different scale in dermatologist Magnus Lynch’s test of 200 biopsied lesions.
A specialized CNN reached 62% diagnostic accuracy against 42% for each of two general models, but the small sample and harder-than-usual cases limit the conclusion.
Cerebras, meanwhile, explains how spreading Kimi K3 across 59 wafers keeps weights on-chip, while DetectorScope tries to expose the features behind a text detector’s decision.
Across infrastructure, medicine and evaluation, the useful advance is not merely a larger model or a cheaper token: it is evidence about how a system behaves, where it fails and whether its benefits survive contact with real work.
windflash.us/daily-report/en…
The strongest audience signals today are specific rather than abstract. Players are not merely asking for more Mass Effect in the sense of another logo on a release calendar. Their petition describes staffing, funding and time, the conditions needed to make the kind of RPG they value. Phantom Blade Zero's estimated pre-orders also show that a new action game can build substantial demand before launch. Yet the first reviews demonstrate why demand is only the opening test: spectacular animation and fluid fighting still need balanced bosses, reliable performance and clear progression to sustain that early confidence.Try New Games
The same distinction applies to games with longer lives. Peak's enlarged map rotation offers replayability without promising endless development, and Diablo-sized expectations are not required for every successful title. Knightmare's modest, free, hand-drawn adventure is equally instructive. It has a clear audience and a bounded creative promise. These projects suggest a healthier measure of ambition than budget or content volume alone: know what players came for, finish that experience properly, and be honest about what happens after release. That is my editorial reading of today's news, not a prediction of commercial results.
best-games.io/game-daily/202…
OpenAI says its models found a serious PHP bug, and PHP runs something like 70% of the web. My sites run on Ghost and Next.js, so I'm probably fine. Still ran the check, because "probably" is how you get hacked. If you have an old WordPress site sitting around, today's a good day.
This Steam Next Fest has 8,011 demos, which is almost double the amount in June.
More and more outsiders like me are using AI for indie game dev. Making a playable game isn't hard anymore; the hard part is getting strangers to actually stop and give it a try.，难的是让更多陌生人停下来，真的点开玩一玩。
nitter.cf/Pirat_Nation/status/21…
Steam Next Fest is facing huge backlash over the rise of AI-slop and low-effort games.
The October event has 8,011 demos, nearly twice as many as June., developers say the flood of low-quality “AI slop” is making it harder for real indie games worth to play
Back in June, roughly 1 in 5 games at Steam Next Fest disclosed using AI. Now, just a few months later, that number has jumped to around 1 in 3.
Some developers have even avoided public demos because they fear their games could be copied.
DeepSeek has open-sourced their inference framework, DeepInfer.
An inference framework is the software used to deploy trained models and actually handle user requests.
DeepInfer is the exact same stack DeepSeek uses to run their DeepSeek-V4.1-Flash model online. It’s under the Apache-2.0 license, so it's free for commercial use.
It supports both NVIDIA SM100 series GPUs and Huawei Ascend 950 chips, featuring several acceleration technologies:
Low-level compute kernels optimized for specific chips,
Model sharding to run a single model across multiple chips simultaneously,
Speculative decoding—basically guessing the next few words and validating them all at once to speed up generation.
However, they only open-sourced the parts not used for their own public online services. Several key features required for large-scale online serving were held back:
Manage the cache for intermediate model computation results,
Split the user input reading and answer generation steps across different machines,
Distribute requests among multiple machines.
This project hit the front page of Hacker News and got 2,027 points.，得了 2,027 票。
github.com/deepseek-ai/DeepI…
Read together: DeepSeek V4.1-Flash is now the top open model on Code Arena WebDev, and per WSJ, DHS just ordered agencies to remove DeepSeek and products built on it. For builders, model choice is becoming a compliance decision, not a benchmark one.
nitter.cf/rodalmighty/status