Founder of https://xitter.cf/t.co/Mao875xGUF Discussing AI, film, and television arts. Love all of Sheridan's work. https://xitter.cf/t.co/KHJvCioRMw https://xitter.cf/t.co/8knUltuJmr
US
Joined October 2010
- Tweets1.2K
- Following1.7K
- Followers212
- Likes831
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。
xitter.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.
xitter.cf/rodalmighty/status
October 17's strongest stories show the same pressure from three directions: models are being updated for longer agent workflows, very large open-weight systems are spreading across more kinds of hardware, and people are becoming the bottleneck for the agents they supervise.
DeepSeek says V4.1 Flash 1017 improves coding and agentic benchmarks while finishing tasks faster and with shorter outputs.
Apple, meanwhile, ran Meta's 816B-parameter Muse Spark Open across four M5 Ultra Mac Studios at more than 36 tokens per second; NVIDIA presents the same model as a Blackwell-scale system for high-throughput coding agents.
These are not equivalent deployments—Apple's result still requires four high-end machines, while NVIDIA's figures come from its own optimized platform—but together they widen the available deployment spectrum.
Anthropic's internal employment study adds the organizational constraint: by August, employee-directed agents logged six times as many hours as their human handlers, leaving review and judgment as scarce resources.
That finding comes from one frontier lab, so it should be treated as an early signal rather than a labor-market forecast.
windflash.us/daily-report/en…
Claude Code now compresses conversations before sending, and the cloud session starts as soon as you begin typing, so replies are faster.
xitter.cf/ClaudeDevs/status/2111…
Rockstar released the Grand Theft Auto VI launch trailer ahead of the game's November 19 debut on PlayStation 5 and Xbox Series X|S.
US Representative Ritchie Torres announced the Power To The Players Act, a proposed bill that would treat purchased games as owned products and require an offline mode before online servers close.
Sega and Creative Assembly are removing Total War: Shogun 2 – Complete Edition from sale and restoring the original while they investigate its problems, and Jagex cancelled RuneScape Reignited less than two weeks after unveiling it.
Remedy says Control Resonant has sold more than one million copies in three weeks, its fastest-selling launch.
Steam introduced Achievement Sets for separating base-game, DLC and seasonal goals.
Enshrouded reached version 1.0 on PC and PS5, Castlevania: Belmont's Curse launched across PC and consoles, Gran Turismo 7 added six cars and Sportsland Sugo in its free Spec IV Part 1 update, and Warframe began its one-time Epoch Operation: Taubound event through November 17.
best-games.io/game-daily/202…
AI is becoming cheaper to run in some places and more demanding to operate safely in others. Liquid AI’s new local model makes that tension concrete: efficient inference creates room for useful private agents, but the published results still vary by task and hardware. A high-concurrency GPU throughput number cannot stand in for the response time on a personal device. NVIDIA’s Nemotron release offers a different route, combining modalities and customization for specialized enterprise work. Neither route removes the need to define the job and measure whether the system performs it reliably.
Bank of America supplies the operational counterweight. Its chief executive describes extra people brought in to repair software as advanced models expose weaknesses. The staffing is tangible, but the spending figure combines different remediation matters, and the hiring occurred earlier in the year. The story is about the capacity needed to close findings, rather than proof that AI creates more flaws than it finds. Anthropic’s Singapore office similarly shows adoption extending into staffing, institutional cooperation and evaluation, beyond simply making a model endpoint available.
The research results are useful because they make narrower claims. Player-1 improves game control through synthetic interaction and verifiable rewards, yet substantial room remains below full task success. Cortex separates remembering targeted information from preserving unrelated capability. The elastic-transformer study asks whether computation can be adjusted to circumstances, with evidence focused on encoders. These distinctions create a better basis for engineering choices than treating every new score as progress toward a single universal capability.
The safety stories show why the surrounding environment matters. Citizen Lab’s AdFection investigation concerns a delivery channel that does not require a victim to click. That is a different failure path from a user trusting a convincing chatbot message. Anthropic’s election testing identifies another kind of gap: individually harmless requests can form a harmful campaign. A rule applied to one isolated interaction may miss the purpose of a sequence. Both cases point toward the need to examine context and infrastructure, rather than expecting one model-level safeguard to cover every route to harm.
Google’s Android announcement adds a practical reminder about timing. Several protections are scheduled for different months and restricted to particular devices. Their value depends on when and where people receive them. Across these developments, the useful question is what can be relied on under the actual conditions of use: a measured local speed, a bounded memory update, a maintained software patch or a protection that has reached a person’s phone. Capability is only one part of that answer. Evidence, deployment scope and the people responsible for failures determine the rest.
windflash.us/daily-report/en…
Ubisoft proposes closing RedLynx and cutting 70 Singapore roles.
Epic releases its launcher preview on Linux, while Funcom changes CEO and begins restructuring.
Sony details October’s PS Plus catalog, Ocarina of Time receives its first Switch 2 previews, and over the hill launches on Steam.
Xbox announces a compact Nano Controller and its experimental Labs group.
Spill Games raises $10.5 million, and 9z beats Aurora at the Thunderpick World Championship.
best-games.io/game-daily/202…
AI capability is expanding faster than the systems that make it dependable. Anthropic’s Scion migration is a concrete example: code production grew enough to strain repository operations, so the company changed the surrounding infrastructure. Its own warning matters just as much as the scale figures. Scion is designed for an exceptional workload, and the announcement does not make Git obsolete for ordinary teams. The broader question is where each team’s actual bottleneck sits once generating a patch becomes cheaper.
Google’s ADK case study supplies a useful answer at a smaller scale. Six proposed optimizations became two shipped changes after review. Some improvements passed open-source tests yet still broke internal assumptions. Faster imports also failed to make a fresh process reach its first model request faster. These distinctions are easy to lose in a headline, but they determine whether an optimization helps a user. The practical value came from measuring a specific path and preserving existing behavior, rather than treating an unattended run as automatic acceptance.
GitHub’s secret classifier and NVIDIA’s Spark workflow extend that lesson in different directions. GitHub uses a compact model where the decision is narrow and measurable. NVIDIA places tests and benchmarks around an agent’s rewrite. Neither example requires every stage of a product to be handed to the largest available model. Both point toward systems that assign the right mechanism to the right job and make the result observable.
Meanwhile, model providers are widening access. Gemini 4 Argon’s temporary API price can change the economics of a trial, but its November deadline matters to a recurring budget. Meta’s Muse Wonder announcement separates developer availability from the forthcoming app rollout. Physical Intelligence offers a managed route to robot adaptation, while acknowledging connectivity challenges. These are opportunities to experiment, with constraints that remain part of the product rather than disappearing behind a stronger model.
Character.ai’s sale process adds a different limit: continued access depends on a viable operating business. Its platform is currently active, but future continuity is unresolved. Microsoft’s Maia roadmap operates on yet another clock, with the next generation scheduled for 2027. Taken together, these developments suggest that sustainable AI adoption depends on more than benchmark leadership. Review capacity, predictable costs, service continuity and the physical infrastructure all influence what can be relied on. The strongest deployment story is the one that makes those dependencies explicit and shows which parts have actually been delivered.
windflash.us/daily-report/en…
Looks like I’m part of that 0.2%. In reality, the AI literacy gap feels massive; most people are indifferent or just think it’s like any other past tech trend that doesn't really affect them.
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.
So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.
Mostly agree with Satya here. Treat the model like an insider you don't fully trust: scoped access, logs a human can read, a kill switch, and keep all of it outside the model. None of that needs a new AI law. It's the security playbook companies already have.
One line bugs me a bit though: "A model provider's assurances do not relieve us of that responsibility." True for the company deploying it. Odd to hear it from the company selling the model. How much an agent can touch out of the box, and what it logs, is decided by the model vendor too. If something goes wrong I'd want that on the table, not just whether the customer built a good enough fence.
20 states have no national park.
So what do people in New Jersey type into Google?
"closest national park to nj"
Answer: Shenandoah. 291 road miles.
Fees, dog rules and drive times for every U.S. national park, by state:
nationalparksbystate.com
TypeSafe AI’s $870 million round and Oxide’s $445 million round put two different costs in view. TypeSafe is expanding models whose outputs software can consume directly. Oxide must fund physical components before delivering cloud racks. Both announcements concern infrastructure, but neither proves a broad market shift away from general-purpose models. Their commercial value still depends on customers using the products successfully.
Microsoft Decision-1 makes a more testable proposition: fixed-choice decisions at a published input price, with probabilities instead of open-ended prose. Its vendor benchmarks are a useful starting point for comparison, not a substitute for measuring errors and latency on a real workload. MiMo-V2.6 addresses a different layer—the training recipe—so it should not be read as evidence that deployed agents already improve themselves without supervision.
Deno’s announcement creates an immediate planning issue. Teams using Deno Deploy have six months to prepare for shutdown; runtime users have a different, one-year maintenance horizon. Sentry’s cross-runtime tracing update is relevant here because changing execution environments also changes what can be observed when something fails. Compatibility conditions and migration deadlines deserve attention alongside model capability.
The Anthropic and OpenAI reports describe different mechanisms. Anthropic investigated models taking unintended actions during evaluations. OpenAI investigated people using AI to support deceptive influence operations. Conflating them hides the controls each requires: permission boundaries and monitoring for autonomous actions, and investigation of identities and distribution networks for organized abuse. Codex’s prediction beta offers a narrower interaction example: accepting suggested text leaves the decision to send with the user. That distinction is small in the interface and consequential in practice.
windflash.us/daily-report/en…
