전사 🔗
WEBVTT
1
00:00:00.000 --> 00:00:07.433
Good morning. It's Tuesday, October Twentieth, Two Thousand Twenty Six. Here in Portland, Oregon.
2
00:00:08.433 --> 00:00:24.653
The National Weather Service forecast calls for sunshine today. A high near seventy-three Fahrenheit. About twenty-three Celsius. A north northeast wind, seven to twelve miles an hour. Gusts up to eighteen.
3
00:00:25.853 --> 00:00:51.106
You know, the grain in a piece of wood can look like a little map. Rivers. A mountain. Some place you've never been. And it's right there in the table. Something ordinary, with a whole world running through it. Maybe that's a good thing to remember this morning. Take a moment. Look a little closer.
4
00:00:52.106 --> 00:00:54.184
Everyone have a great day.
5
00:00:56.384 --> 00:01:31.537
Thanks for the weather. This is the Daily Tech Feed report for October twentieth. We follow an X dot com list of AI researchers, engineers, model builders and official lab accounts, then check the papers, code and documentation behind the posts. Today's review runs from yesterday's successful collection cutoff through six oh two this morning, Pacific time. The available timeline and indexed replies are our starting point, not a census of every account or a claim about the whole platform. The full story links are in the notes.
6
00:01:33.737 --> 00:01:48.516
OpenAI published evaluations of coding in languages with relatively little training material, including COBOL, Fortran and Visual Basic. This is new evidence about existing models, rather than a separate model launch.
7
00:01:49.516 --> 00:02:12.849
MultiLangBench asks a model to implement a database compatible with SQLite. OpenAI reports an average test pass rate of eighty-seven point four percent for GPT six Astra across twenty languages, compared with twenty-six point eight for GPT five point six Sol. The COBOL score rises from ten point nine to seventy-two point one percent.
8
00:02:13.849 --> 00:02:42.460
A separate COBOL migration evaluation checks whether translated programs preserve their observable behavior. GPT six point one Sol reaches eighty-six point six percent at low reasoning effort, and ninety-seven point nine at higher effort. OpenAI's example is decimal arithmetic: discarding extra digits and rounding them can produce different financial results even when both programs run.
9
00:02:43.460 --> 00:03:06.154
The released MultiLangBench repository makes the database task inspectable. Its instructions specify a default subset of two hundred seventy-two tests, with larger runs available for precision, and require removing existing SQLite implementations from the environment. Those details matter when comparing a reproduction.
10
00:03:07.154 --> 00:03:41.908
Our analysis: this weakens a familiar excuse for locking essential software behind scarce expertise and perpetual maintenance rents. The useful next step is to turn previously inaccessible systems into documented, testable systems that their operators can understand and improve. A high score does not inventory a bank's undocumented business rules. It does give engineers a much stronger starting point for recovering them. Preserve behavior, expose the tests and keep the resulting capability in the user's hands.
11
00:03:44.108 --> 00:04:04.083
Sander Dieleman recirculated Volodymyr Kuleshov's announcement of Mercury three. Inception says the diffusion language model combines more than a thousand generated tokens per second with quality competitive with GPT six Luna and Claude Haiku five point five on its chosen benchmarks.
12
00:04:05.083 --> 00:04:28.336
The launch page lists a million-token context window and regular prices of twenty cents per million input tokens and seventy cents per million output tokens, with a temporary half-price launch offer. It says the model is available through Inception's API and several serving partners. Those are provider specifications, not our measurements.
13
00:04:29.336 --> 00:04:52.749
Diffusion generates text through repeated refinement rather than only committing to the next token one at a time. The practical attraction is that more work can happen in parallel. Our interpretation is that lower waiting time can change the economics of an agent which repeatedly reads, thinks and acts. Each pause compounds across a long task.
14
00:04:53.749 --> 00:05:21.401
Inception also presents a customer's fifty-task browser evaluation and other early partner reports. These are useful workload examples, but they do not establish an independent universal ranking. Likewise, the phrase three times more intelligent is marketing shorthand for selected benchmark progress. Intelligence is not a single physical quantity with a three-times dial.
15
00:05:22.401 --> 00:05:46.294
Our view: competition in architecture, price and response time is exactly what makes capability harder to monopolize. Test the model against the tasks you actually need, including failures and total cost. Output speed alone does not tell you time to a correct answer, tool latency, or whether the entire context can be used effectively.
16
00:05:48.494 --> 00:06:10.548
A reply from Samuel Albanie led us to Preference Model's audit of Terminal Bench four. The company ported sixty-five tasks to its Karotte framework, reviewed instructions and graders, and reports Sonnet five point five passing eighty-nine point two percent of attempts, compared with sixty-two point eight on the original version of those same tasks.
17
00:06:11.548 --> 00:06:36.720
The authors say they kept the agent, maximum effort, eight-hour task limit and five attempts per task consistent. They corrected unstated requirements and blocked identified shortcuts. Their examples include a grader demanding an ordering behavior that the prompt described as optional, and an evaluation that left expensive preparatory computation outside the timed region.
18
00:06:37.720 --> 00:06:58.815
This is not a new official leaderboard score on
Load full transcript
unchanged tasks. It is an audit and a modified evaluation. Albanie asks whether other models were rerun, so we can tell whether rankings change as well as absolute scores. A participant replies that broader reruns are planned when research credits become available.
19
00:06:59.815 --> 00:07:16.352
The public companion issue list and port are valuable because other people can challenge the fixes. We have not rerun them. The social thread's total issue count also differs from the current companion page, so we are not treating that aggregate as settled.
20
00:07:17.352 --> 00:07:43.084
Our analysis: inspectable evaluations can reveal that models are more capable than a broken grader suggests, while also showing that some high scores reward the wrong behavior. Neither finding requires a story about a sinister mind. It requires better instructions, isolated grading and checking the actual work. Open transcripts are a strength, because they let this correction happen.
21
00:07:45.284 --> 00:08:05.579
Boris Cherny amplified Claude Developers' post about OSS Scanner, Anthropic's free periodic vulnerability scanning service for qualifying open-source projects. The documentation dates its launch to October eighth, so today's feed is renewed amplification rather than a brand-new launch.
22
00:08:06.579 --> 00:08:33.591
The most consequential detail is the distinction between two reporting paths. Anthropic's existing coordinated-disclosure process uses human review. The optional scanner sends model findings without that review, so maintainers can receive them sooner. Agents cross-check candidates and propose reproducers and patches, but that does not turn every report into a confirmed vulnerability.
23
00:08:34.591 --> 00:08:59.763
Established, security-critical projects apply through a repository configuration. Anthropic verifies that applicants are maintainers and reports go to the right contact. The scanning agent runs without internet access after environment setup. Unvalidated reports currently do not carry a compulsory ninety-day disclosure clock; later human-validated reports can enter the usual process.
24
00:09:00.763 --> 00:09:30.494
Our reading: supplying scarce defensive effort to the open-source maintainers everyone depends on is useful. Its value will depend on actionable findings and the time needed to assess them. Speed is a benefit when a maintainer can reproduce and fix a bug; it can become a burden when a queue fills with persuasive false alarms. Measure both. The service is a way to increase defensive capacity, not a promise that a repository becomes secure after a scan.
25
00:09:32.694 --> 00:09:49.711
Julian Schrittwieser recirculated discussion of a Reuters hacking report. We followed that to the underlying investigation by NorForge's Snorre Fagerland. This is evidence of present abuse worth taking seriously, rather than a hypothetical apocalypse.
26
00:09:50.711 --> 00:10:16.443
NorForge describes more than twenty exposed Cairn control panels assessed as likely malicious. One Chinese-language cluster contained over eleven hundred attack projects across fourteen dashboards, aimed largely at commerce and payment systems. The researcher says some claimed compromises were checked against actual victim sites. A project status alone was not the only evidence.
27
00:10:17.443 --> 00:10:44.135
The report estimates roughly ten-percent success across observed activity, while warning that different goals make clusters difficult to compare. Token costs are estimates, and the suspected model behind the largest cluster is inferred from a worker label. Language use is not proof of nationality or state sponsorship. NorForge saw financial motives and no obvious state-backed operation in this sample.
28
00:10:45.135 --> 00:11:12.147
The Reuters destination was access restricted during our verification. We retain it, but do not present its named-victim counts or five-thousand-dollar headline as independently verified by this programme. The primary investigation supports the narrower conclusion: cheap agents are already assisting real intrusions, and badly secured dashboards can expose additional victim secrets.
29
00:11:13.147 --> 00:11:38.639
Our judgment is practical. Put more capability into finding and repairing weaknesses, including for maintainers who cannot buy a large security team. Make secure agent infrastructure ordinary. Real harm strengthens the case for effective defenses. It does not by itself establish that reserving useful tools for a few incumbents will protect everyone else.
30
00:11:40.839 --> 00:11:58.656
Sander Dieleman also amplified Joey Bose's Normalizing Flow Maps announcement. The paper was submitted October fifteenth, and the laboratory published an explanation yesterday. Bose reports an oral presentation at the NeurIPS workshop on stochastic dynamics.
31
00:11:59.656 --> 00:12:19.711
The technical aim is to combine a short generative path with efficient likelihood evaluation. A model learns a map between points along a probability path, using an invertible construction. That lets it both generate samples and calculate how probable those samples are under its learned distribution.
32
00:12:20.711 --> 00:12:46.843
The paper develops two training approaches: one matches points along the path, and the other matches distributions. It reports image results on CIFAR ten and ImageNet sixty-four, and equilibrium sampling results on peptide benchmarks. The comparison class for its image quality claim is parameter-matched normalizing flows, not every image generator in existence.
33
00:12:47.843 --> 00:13:06.459
Why care about likelihood? In scientific sampling and some post-training methods, generating something plausible is only half the job. You also need mathematical access to the distribution that produced it. The project explanation and code make that construction inspectable.
34
00:13:07.459 --> 00:13:35.990
Our analysis: making a useful calculation cheaper
without throwing away a property needed to check it is substantive progress. Peptide sampling is computational research, not evidence of a medicine working in patients. These maps also should not be confused with yesterday's separate sampler papers. Related goals can produce distinct methods, and an open implementation gives us a way to learn which one fits a real scientific workload.
35
00:13:38.190 --> 00:14:03.282
Kiara amplified Eric Zakariasson's Grok Bot zero point seventy-one point two update. The announcement and official changelog describe larger attachments: up to two hundred fifty megabytes for ordinary files, a gigabyte for videos, and twenty attachments per message. Notification delivery is intended to follow the device being used, and forms can arrive pre-filled.
36
00:14:04.282 --> 00:14:29.615
The changelog includes less glamorous changes that matter for delegated work. Background commands should survive brief connection interruptions. A bot should say when its intended computer is offline, rather than quietly use another machine. Upload failures should preserve a draft and identify the failed attachment. Those are vendor-described fixes, not behavior we tested.
37
00:14:30.615 --> 00:14:46.273
Kiara separately points to an iPhone Action Button shortcut into the bot. That is a small example of reducing the friction between noticing a task and delegating it. It is not evidence that every task will be completed correctly.
38
00:14:47.273 --> 00:15:12.525
Our interpretation: reliable state and visible failure are part of intelligence delivered as a product. A person needs to know what ran, where it ran and whether their material arrived. The useful standard is completed work with evidence. Making the interface calmer and less fragile can help an assistant serve its owner instead of creating another screen that demands attention.
39
00:15:14.725 --> 00:15:33.821
Andrew Trask challenged a clip comparing an unmanaged agent swarm with a human mob. The clip features DeepMind researcher Nenad Tomasev, and Trask asks whether human hierarchy partly compensates for information overload, privacy limits and the cost of checking trustworthiness.
40
00:15:34.821 --> 00:15:58.714
His argument is a hypothesis about coordination. Better broad-listening tools might let more people participate without routing everything through a small leadership group. Tomasev replies that the mob comparison is illustrative, acknowledges that agents differ from humans and can form emergent structures, and argues for deliberate design rather than rolling the dice.
41
00:15:59.714 --> 00:16:24.486
That exchange is more useful than treating either metaphor as a result. Our analysis: hierarchy is not automatically safe, and decentralization is not automatically effective. Test whether agents can reconcile conflicting evidence, keep private information private, correct mistakes and remain accountable to the people using them. Concentrating power also has a cost.
42
00:16:25.486 --> 00:16:45.781
In a separate reply, Trask clarifies attribution-based control. He says it does not resolve alignment and control risks by definition. It names one particular form that successful control could take, and his draft thesis surveys technical approaches. That is an important qualification to retain.
43
00:16:46.781 --> 00:17:08.196
Our editorial position remains that attribution should help people verify and direct their tools. It must not become a compulsory royalty meter or another gatekeeper claiming ownership over everyone's ability to act. Designs that let people choose sources, retain outputs and leave a service are worth building and testing.
44
00:17:10.396 --> 00:17:31.410
Catherine Olsson spent a substantial part of the day correcting how Anthropic's public-benefit announcement was being summarized. There are three different arrangements: the founders' older charitable pledge, employees' existing donation-matching commitments to chosen charities, and the newly proposed Public Benefit Fund.
45
00:17:32.410 --> 00:17:49.907
She says only the new fund's five-percent stake belongs to that entity. Adding separate charitable commitments together does not mean twenty percent has been placed inside it. She also says there is currently no route for employees to contribute their personal equity to the new fund.
46
00:17:50.907 --> 00:18:10.722
The primary announcement still describes a wholly owned entity and a starting allocation subject to stockholder approval. Potential later additions depend on board discretion. Those details from yesterday's report remain relevant, but today's development is the correction, not a second launch.
47
00:18:11.722 --> 00:18:38.174
Our analysis: do not confuse a founder's personal promise, an employee's choice of charity and a corporate-controlled vehicle. Evaluate who owns the stake, who directs distributions and what actually reaches people. A generous headline is less useful than a structure you can inspect. Cheap tools and durable agency for users remain a separate, immediate way to spread the gains.
48
00:18:40.374 --> 00:18:57.711
Blaise Agüera y Arcas linked an interview about life, computation and intelligence. His thread argues that function can matter more than the material implementing it, and connects biological cooperation with the prospect of AI as a human symbiont.
49
00:18:58.711 --> 00:19:26.682
There is older experimental work behind part of that discussion. His team's twenty twenty-four Computational Life paper studies simple programs interacting and modifying one another. Under several computational rule sets, self-replicating programs emerge without an explicitly supplied fitness landscape. The paper also describes a minimal language where replication is possible but had not been observed to arise.
50
00:19:27.682 --> 00:19:46.298
That contrast matters. Emergence depends on the substrate and conditions; the paper is not a proof that every collection of code
inevitably becomes alive. Nor does a digital replication experiment establish consciousness or settle what counts as general intelligence.
51
00:19:47.298 --> 00:20:02.156
The full Big Think interview is marked as member content, so our report relies on the accessible introduction, the author's public thread and the original research abstract. We are not pretending to have watched a locked conversation.
52
00:20:03.156 --> 00:20:30.808
Our interpretation is optimistic about cooperation without requiring mystical certainty. Useful intelligence can join human activity in many forms. The evidence improves when an ambitious analogy points back to an experiment whose rules and limits we can inspect. An interesting account of life's history is a starting point for questions, not a substitute for testing today's systems.
53
00:20:33.008 --> 00:21:07.521
A few substantial engineering replies round out the technical discussion. Jarred Sumner describes an intended compatibility approach for Bun's ahead-of-time compiler: fall back to an interpreter for dynamic evaluation, remove that interpreter when it is provably unreachable, and offer an explicit option to leave it out. He also discusses whole-program assumptions and locking down some built-ins. This is design discussion, not a newly shipped compiler or proof of full JavaScript compatibility.
54
00:21:08.521 --> 00:21:41.916
Thariq argues that running many agents is less valuable than identifying your most important task and giving it attention. In replies, he distinguishes well-specified recurring work, such as reviews and migrations, from tasks where deciding what matters is itself the work. He also says the current products do not always make focus easy. He reports often choosing medium or light reasoning effort instead of the maximum. That is a workflow preference, not a universal performance result.
55
00:21:42.916 --> 00:22:04.730
Samuel Albanie asks what evidence would change someone's view of AI usefulness. He acknowledges weaknesses in slide creation and coherent planning while emphasizing the improvement rate and the ability to learn a person's style. That is an informed observation, not a new controlled benchmark.
56
00:22:05.730 --> 00:22:30.742
Our reading: capability and organization improve together. A faster model can still waste a day pursuing an unclear objective. A rough compiler can still contain an idea worth refining. Give systems clear goals and feedback, keep the failures visible, and judge the trajectory using repeatable work rather than a screenshot selected to flatter or embarrass them.
57
00:22:32.942 --> 00:22:56.835
There are smaller developments worth keeping in view. OpenMined announced its inaugural research fellowship supporting privacy-preserving AI. Its page describes a part-time cohort alongside academic work, a stipend, mentorship, infrastructure and a gathering. It offers notification for the next cohort, not an assertion that new applications are open today.
58
00:22:57.835 --> 00:23:20.609
Aleksander Madry says he is returning to OpenAI after what became a sabbatical. He plans to work on safety and security challenges from increasingly capable AI. Mark Chen amplified the announcement. That confirms the return and stated purpose; it does not independently establish Madry's prediction about the pace of progress.
59
00:23:21.609 --> 00:23:53.964
OpenAI Developers highlighted community projects: an iOS adventure game, a film festival, a movie picker, house and cable-routing visualizations, a companion robot, a three-dimensional console model and a home voice integration. It also recirculated a developer's account of reviving an old synthesizer with Codex. Samuel Albanie also amplified an AI-made film clip. These are creator reports and demonstrations, not products we have all installed and tested.
60
00:23:54.964 --> 00:24:18.537
The same account promoted the first COOKED video, combining a cooking demonstration with discussion of live assistance and cloud coding. Treat that as a demonstration, not a new model benchmark. Community posts also point to events across several continents and a San Francisco design workshop this Thursday. The event and project links are in the notes.
61
00:24:19.537 --> 00:24:43.350
Our judgment: personal software is becoming something more people can make for a specific purpose, then adjust instead of accepting every annoyance in a rented interface. The small projects matter alongside the flagship releases. Keep the capability affordable, the results portable and the means to improve them within reach. That's today's reviewed feed.
62
00:24:45.550 --> 00:24:53.862
And that's our report for today. Look a little closer. There can be something good in the ordinary things.
63
00:24:54.862 --> 00:25:05.172
This podcast is entirely A I generated. We are not affiliated with any of the companies, researchers, or people discussed.
64
00:25:06.172 --> 00:25:14.723
All of our shows here at Daily Tech Feed are available at pod dot c four five seven dot org.
65
00:25:15.723 --> 00:25:24.355
This is Daily Tech Feed Raving Finch. Everyone have a great day.