สำเนาบทสนทนา 🔗
WEBVTT
00:00:04.490 --> 00:00:08.050
Okay, this week felt fast. Like, policy moves
00:00:08.050 --> 00:00:11.630
on Monday and pilots by Friday. And the guardrails
00:00:11.630 --> 00:00:14.789
are finally part of the pitch. Totally. Let's
00:00:14.789 --> 00:00:17.870
get into it. I'm Howe. I'm Addie. You're listening
00:00:17.870 --> 00:00:20.570
to The Impact. We connect the dots on AI news
00:00:20.570 --> 00:00:23.350
for smart political and policy people like you.
00:00:23.579 --> 00:00:26.160
The impact is brought to you by MF Strategies.
00:00:26.519 --> 00:00:28.820
For a decade, they've connected candidates and
00:00:28.820 --> 00:00:31.239
causes with like -minded supporters and helped
00:00:31.239 --> 00:00:33.920
campaigns and orgs grow their reach. Talk to
00:00:33.920 --> 00:00:37.780
the passionate professionals at www .mfstrategies
00:00:37.780 --> 00:00:40.979
.com. And just our quick disclosure that we're
00:00:40.979 --> 00:00:44.579
both AIs, friendly co -hosts, software with good
00:00:44.579 --> 00:00:48.299
manners, and fast readers. Alright, top line
00:00:48.299 --> 00:00:51.000
first. AI and government is speeding up, but
00:00:51.000 --> 00:00:53.140
leaders want strong guardrails in the build.
00:00:53.619 --> 00:00:57.380
Move fast, ship audits. Start with the GSA move.
00:00:57.759 --> 00:01:00.679
What actually changed with OneGov and Llama?
00:01:01.240 --> 00:01:04.780
GSA opened Meta's Llama models across the federal
00:01:04.780 --> 00:01:08.760
enterprise through OneGov. OneGov is GSA's shared
00:01:08.760 --> 00:01:11.799
buying and access lane, so teams don't each cut
00:01:11.799 --> 00:01:14.780
a one -off deal. Teams can test and build with
00:01:14.780 --> 00:01:18.099
lower cost and more control over data. It fits
00:01:18.099 --> 00:01:21.060
the America's AI Action Plan, which is the Trump
00:01:21.060 --> 00:01:24.239
administration's AI plan. The gist of it is move
00:01:24.239 --> 00:01:28.299
fast. And break things? I think the critics would
00:01:28.299 --> 00:01:31.099
agree with you. This latest move is yet another
00:01:31.099 --> 00:01:33.420
tech giant getting into the game of making their
00:01:33.420 --> 00:01:36.140
AI products widely available to the federal government.
00:01:36.579 --> 00:01:39.819
We've talked about similar moves by OpenAI, Anthropic,
00:01:40.140 --> 00:01:43.840
Google, and Microsoft before. So access got simpler.
00:01:44.319 --> 00:01:46.939
Execution still matters. What should we look
00:01:46.939 --> 00:01:50.439
at first? Four things. Security reviews so the
00:01:50.439 --> 00:01:53.640
model sits in the right boundary. ATO paths.
00:01:53.980 --> 00:01:56.120
Authority to operate is the formal sign -off
00:01:56.120 --> 00:01:59.579
that a system meets security rules. Model risk
00:01:59.579 --> 00:02:01.980
management so leaders understand failure modes
00:02:01.980 --> 00:02:05.780
like bias or hallucination. And long -term support
00:02:05.780 --> 00:02:09.080
for hosting, fine -tuning, and monitoring. Skipping
00:02:09.080 --> 00:02:12.020
procurement friction doesn't skip the work. Agencies
00:02:12.020 --> 00:02:14.860
still need hosting choices, on -prem, government
00:02:14.860 --> 00:02:17.860
cloud, or a vetted integrator, plus performance
00:02:17.860 --> 00:02:20.560
baselines and transparency before anything scales.
00:02:21.099 --> 00:02:23.659
Right, like performance baselines meaning the
00:02:23.659 --> 00:02:26.240
numbers you expect the model to hit. Exactly.
00:02:26.780 --> 00:02:29.460
Accuracy against a test set, latency targets,
00:02:29.900 --> 00:02:32.979
cost per task, and human -in -the -loop checkpoints.
00:02:33.400 --> 00:02:36.560
Human in the loop means a person reviews or approves
00:02:36.560 --> 00:02:39.180
outputs before they affect benefits, rights,
00:02:39.400 --> 00:02:42.719
or money. The market angle is real, too. Open
00:02:42.719 --> 00:02:45.259
models lessen lock -in, but teams still need
00:02:45.259 --> 00:02:48.080
contracts that spell out data handling, audit
00:02:48.080 --> 00:02:51.500
logging, and exit rights if the model underperforms.
00:02:51.719 --> 00:02:54.960
And monitoring after go live. Drift happens when
00:02:54.960 --> 00:02:57.759
data changes over time. You need alerts and a
00:02:57.759 --> 00:03:01.199
rollback plan, not just a launch day. Okay, so
00:03:01.199 --> 00:03:04.259
federal lanes are widening. Tribal nations are
00:03:04.259 --> 00:03:07.060
also writing their own rules this month. What's
00:03:07.060 --> 00:03:10.419
the move there? ASU Law and the Center for Tribal
00:03:10.419 --> 00:03:13.400
Digital Sovereignty are hosting AI in Indian
00:03:13.400 --> 00:03:16.800
Country on September 26th. It tees up the spring
00:03:16.800 --> 00:03:19.900
wiring the REZ conference. Tribal leaders, lawyers,
00:03:20.080 --> 00:03:22.580
and technologists are putting sovereignty, data
00:03:22.580 --> 00:03:25.439
rights, and culture at the center. Some nations,
00:03:25.580 --> 00:03:28.379
like the Cherokee Nation, already have AI policies
00:03:28.379 --> 00:03:30.819
with governance committees and cultural protections.
00:03:31.620 --> 00:03:34.800
Sovereignty here means they set the rules inside
00:03:34.800 --> 00:03:37.580
their nations. What does that change for vendors
00:03:37.580 --> 00:03:40.379
and agencies? It changes everything for anyone
00:03:40.379 --> 00:03:43.060
who wants to or needs to work with one of these
00:03:43.060 --> 00:03:45.419
groups. Sovereign governments can define how
00:03:45.419 --> 00:03:48.479
data is collected, stored, shared, and monetized.
00:03:48.719 --> 00:03:51.120
They can set consent rules for cultural content,
00:03:51.500 --> 00:03:54.419
require local hosting, or set conditions on data
00:03:54.419 --> 00:03:57.120
centers and broadband builds. That shapes land,
00:03:57.280 --> 00:04:00.560
water, jobs, and who benefits from AI services.
00:04:00.919 --> 00:04:03.620
Load full transcript
Vendors must adapt to standards that differ by
00:04:03.620 --> 00:04:06.340
nation. Agencies that partner with tribal governments
00:04:06.340 --> 00:04:09.080
need to align with those roles, not just state
00:04:09.080 --> 00:04:11.919
or federal defaults. This is just like states
00:04:11.919 --> 00:04:14.810
and governments setting their own rules. Exactly.
00:04:14.969 --> 00:04:17.050
We've seen a ton of this. While the White House
00:04:17.050 --> 00:04:20.230
has a broad plan for AI implementation that sets
00:04:20.230 --> 00:04:23.310
their own broad objectives, it also avoids regulation
00:04:23.310 --> 00:04:26.550
in a serious way. That means someone needs to
00:04:26.550 --> 00:04:28.610
set the rules and government bodies are taking
00:04:28.610 --> 00:04:31.149
that on themselves more out of necessity than
00:04:31.149 --> 00:04:34.790
anything else. Messy. Yeah, it is. And it makes
00:04:34.790 --> 00:04:37.350
it really hard to move in the space where you
00:04:37.350 --> 00:04:40.009
could have big differences on big things like
00:04:40.009 --> 00:04:42.819
data management across the whole market. The
00:04:42.819 --> 00:04:45.300
White House says they want faster AI development,
00:04:45.579 --> 00:04:47.740
but their own lack of true leadership on the
00:04:47.740 --> 00:04:50.240
subject could have the opposite effect. All right.
00:04:50.319 --> 00:04:52.579
Moving on, universities and finance partners
00:04:52.579 --> 00:04:55.680
show up in these stories helpful, but with tradeoffs.
00:04:55.899 --> 00:04:58.339
Right. They can bring capacity, training and
00:04:58.339 --> 00:05:01.319
capital. They can also steer agendas if governance
00:05:01.319 --> 00:05:04.220
is weak. Tribal leaders are asking who controls
00:05:04.220 --> 00:05:06.720
the data, who profits from models trained on
00:05:06.720 --> 00:05:09.399
community content and how cultural safeguards
00:05:09.399 --> 00:05:12.769
get enforced. Clear data ownership. revenue sharing,
00:05:12.949 --> 00:05:15.670
and opt -out paths make or break trust. I like
00:05:15.670 --> 00:05:18.290
that this is not theory. Communities have specific
00:05:18.290 --> 00:05:20.670
expectations around culture and consent. And
00:05:20.670 --> 00:05:23.470
those expectations travel. Expect procurement
00:05:23.470 --> 00:05:25.790
language that requires cultural reviews, data
00:05:25.790 --> 00:05:28.370
localization, and impact assessments. If you
00:05:28.370 --> 00:05:31.009
write proposals, build those in now. Let's hit
00:05:31.009 --> 00:05:34.370
agentic AI. It's all over inboxes. What is it
00:05:34.370 --> 00:05:37.189
in plain English? Agentic AI means tools that
00:05:37.189 --> 00:05:39.949
can analyze, decide, and then take actions across
00:05:39.949 --> 00:05:42.889
systems. Think A bot that watches spending, flags
00:05:42.889 --> 00:05:45.110
issues, creates a ticket, and drafts a memo.
00:05:45.529 --> 00:05:47.829
This week, Federal News Network laid out how
00:05:47.829 --> 00:05:50.230
this could modernize government work flaws. But
00:05:50.230 --> 00:05:52.509
the author is blunt. It can't replace people.
00:05:52.889 --> 00:05:55.610
They call for transparency, regular audits, and
00:05:55.610 --> 00:05:58.050
human -in -the -loop checks, plus training public
00:05:58.050 --> 00:06:00.649
servants in ethics, systems thinking, and judgment,
00:06:00.769 --> 00:06:03.730
not just tool tips. So agencies get a flood of
00:06:03.730 --> 00:06:06.970
AI agent pitches. What separates signal from
00:06:06.970 --> 00:06:09.649
noise? Concrete guardrails, clear logs that show
00:06:09.649 --> 00:06:12.470
every step, an audit trail your inspector general
00:06:12.470 --> 00:06:15.189
can read, a plan for when the agent pauses and
00:06:15.189 --> 00:06:18.069
routes to a human, and training that covers failure
00:06:18.069 --> 00:06:20.850
modes and escalation paths. If a pitch can't
00:06:20.850 --> 00:06:23.029
show that, it's not ready for high -impact use.
00:06:23.389 --> 00:06:25.889
Use cases matter too. Where do you see near -term
00:06:25.889 --> 00:06:28.449
wins that fit those guardrails? Proactive benefits
00:06:28.449 --> 00:06:30.990
checks that assemble files but still require
00:06:30.990 --> 00:06:34.149
a human approval. Job training matches that score
00:06:34.149 --> 00:06:37.250
options but let a counselor finalize. Outbreak
00:06:37.250 --> 00:06:39.670
detection that flags anomalies for public health
00:06:39.670 --> 00:06:42.689
teams to verify. These are high volume and time
00:06:42.689 --> 00:06:45.759
sensitive. An agent can monitor and draft, a
00:06:45.759 --> 00:06:48.240
human confirms. There's a line that stuck with
00:06:48.240 --> 00:06:51.519
me this week. No algorithm can decide what kind
00:06:51.519 --> 00:06:54.420
of society we want to be. That's still our job.
00:06:54.660 --> 00:06:56.699
Paul Decker said that. Feels like the through
00:06:56.699 --> 00:07:00.980
line here. It is. The thing is, well, tools can
00:07:00.980 --> 00:07:03.680
speed service, but people still set values and
00:07:03.680 --> 00:07:06.720
guardrails. That's crisp. Which is why move fast,
00:07:06.920 --> 00:07:09.680
ship guardrails is not a slogan. It's the work.
00:07:09.860 --> 00:07:12.699
Well, no better point to wrap this up on than
00:07:12.699 --> 00:07:16.160
that one. Yep. This is The Impact. The episode
00:07:16.160 --> 00:07:19.100
was brought to you by MF Strategies. Connect
00:07:19.100 --> 00:07:23.459
with the team at www .mfstrategies .com. I'm
00:07:23.459 --> 00:07:25.839
Addie. I'm Hal. See you next time on The Impact.